%Version 3.1 December 2024
% See section 11 of the User Manual for version history
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%                                                                 %%
%% Please do not use \input{...} to include other tex files.       %%
%% Submit your LaTeX manuscript as one .tex document.              %%
%%                                                                 %%
%% All additional figures and files should be attached             %%
%% separately and not embedded in the \TeX\ document itself.       %%
%%                                                                 %%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%%\documentclass[referee,sn-basic]{sn-jnl}% referee option is meant for double line spacing

%%=======================================================%%
%% to print line numbers in the margin use lineno option %%
%%=======================================================%%

%%\documentclass[lineno,pdflatex,sn-basic]{sn-jnl}% Basic Springer Nature Reference Style/Chemistry Reference Style

%%=========================================================================================%%
%% the documentclass is set to pdflatex as default. You can delete it if not appropriate.  %%
%%=========================================================================================%%

%%\documentclass[sn-basic]{sn-jnl}% Basic Springer Nature Reference Style/Chemistry Reference Style

%%Note: the following reference styles support Namedate and Numbered referencing. By default the style follows the most common style. To switch between the options you can add or remove “Numbered” in the optional parenthesis. 
%%The option is available for: sn-basic.bst, sn-chicago.bst%  
 
%%\documentclass[pdflatex,sn-nature]{sn-jnl}% Style for submissions to Nature Portfolio journals
%%\documentclass[pdflatex,sn-basic]{sn-jnl}% Basic Springer Nature Reference Style/Chemistry Reference Style
\documentclass[pdflatex,sn-mathphys-num]{sn-jnl}% Math and Physical Sciences Numbered Reference Style
%%\documentclass[pdflatex,sn-mathphys-ay]{sn-jnl}% Math and Physical Sciences Author Year Reference Style
%%\documentclass[pdflatex,sn-aps]{sn-jnl}% American Physical Society (APS) Reference Style
%%\documentclass[pdflatex,sn-vancouver-num]{sn-jnl}% Vancouver Numbered Reference Style
%%\documentclass[pdflatex,sn-vancouver-ay]{sn-jnl}% Vancouver Author Year Reference Style
%%\documentclass[pdflatex,sn-apa]{sn-jnl}% APA Reference Style
%%\documentclass[pdflatex,sn-chicago]{sn-jnl}% Chicago-based Humanities Reference Style

%%%% Standard Packages
%%<additional latex packages if required can be included here>

\usepackage{graphicx}%
\usepackage{multirow}%
\usepackage{amsmath,amssymb,amsfonts}%
\usepackage{amsthm}%
\usepackage{mathrsfs}%
\usepackage[title]{appendix}%
\usepackage{xcolor}%
\usepackage{textcomp}%
\usepackage{manyfoot}%
\usepackage{booktabs}%
\usepackage{algorithm}%
\usepackage{algorithmicx}%
\usepackage{algpseudocode}%
\usepackage{listings}%
%%%%

\usepackage{makecell}
\usepackage{hyperref}
\usepackage{float}
\usepackage{pifont}
\usepackage{pgfplots}

\definecolor{myblue}{RGB}{70, 130, 180}   
\definecolor{myorange}{RGB}{242, 142, 43} 
\definecolor{myred}{RGB}{225, 87, 89} 
\pgfplotsset{compat=1.18}

%%%%%=============================================================================%%%%
%%%%  Remarks: This template is provided to aid authors with the preparation
%%%%  of original research articles intended for submission to journals published 
%%%%  by Springer Nature. The guidance has been prepared in partnership with 
%%%%  production teams to conform to Springer Nature technical requirements. 
%%%%  Editorial and presentation requirements differ among journal portfolios and 
%%%%  research disciplines. You may find sections in this template are irrelevant 
%%%%  to your work and are empowered to omit any such section if allowed by the 
%%%%  journal you intend to submit to. The submission guidelines and policies 
%%%%  of the journal take precedence. A detailed User Manual is available in the 
%%%%  template package for technical guidance.
%%%%%=============================================================================%%%%

%% as per the requirement new theorem styles can be included as shown below
\theoremstyle{thmstyleone}%
\newtheorem{theorem}{Theorem}%  meant for continuous numbers
%%\newtheorem{theorem}{Theorem}[section]% meant for sectionwise numbers
%% optional argument [theorem] produces theorem numbering sequence instead of independent numbers for Proposition
\newtheorem{proposition}[theorem]{Proposition}% 
%%\newtheorem{proposition}{Proposition}% to get separate numbers for theorem and proposition etc.

\theoremstyle{thmstyletwo}%
\newtheorem{example}{Example}%
\newtheorem{remark}{Remark}%

\theoremstyle{thmstylethree}%
\newtheorem{definition}{Definition}%

\raggedbottom
%%\unnumbered% uncomment this for unnumbered level heads

\begin{document}

\title{Multi-view Patch Inference with Frozen DINOv3 for Industrial Anomaly Detection}

%%=============================================================%%
%% GivenName	-> \fnm{Joergen W.}
%% Particle	-> \spfx{van der} -> surname prefix
%% FamilyName	-> \sur{Ploeg}
%% Suffix	-> \sfx{IV}
%% \author*[1,2]{\fnm{Joergen W.} \spfx{van der} \sur{Ploeg} 
%%  \sfx{IV}}\email{iauthor@gmail.com}
%%=============================================================%%

\author[1]{\fnm{Chaoqun} \sur{Wang}}\email{20241081210203@buu.edu.cn}

\author[3]{\fnm{Wenjing} \sur{Zhang}}\email{zhangwenjing242@mails.ucas.ac.cn}

\author[2]{\fnm{Bo} \sur{Qi}}\email{20241083510910@buu.edu.cn}

\author[1]{\fnm{Xiaoyu} \sur{Huang}}\email{20242085410205@buu.edu.cn}
\author*[1]{\fnm{Ning} \sur{He}}\email{xxthening@buu.edu.cn}
\affil*[1]{\orgdiv{College of Smart City}, \orgname{Beijing Union University}, \orgaddress{\city{Beijing}, \country{China}}}

\affil[2]{\orgdiv{College of Robotics}, \orgname{Beijing Union University}, \orgaddress{\city{Beijing}, \country{China}}}

\affil[3]{\orgdiv{School of Engineering Science}, \orgname{University of Chinese Academy of Sciences}, \orgaddress{\city{Beijing}, \postcode{100049}, \country{China}}}

%%==================================%%
%% Sample for unstructured abstract %%
%%==================================%%

\abstract{Industrial anomaly detection is a critical task in automated visual inspection systems, aiming to identify subtle defects without relying on large-scale anomaly annotations. Recently, pretrained vision transformers such as DINO have demonstrated strong feature representation capabilities and have been widely adopted in this domain. However, most existing methods rely on single-grid patch partitioning, which suffers from boundary information loss and limited local context modeling, resulting in degraded performance on small or boundary-located defects.In this paper, we propose a simple yet effective framework based on a frozen DINOv3 backbone for industrial anomaly detection. To address the limitations of conventional patch-based inference, we introduce a Multi-view Patch Inference (MVPI) strategy, which constructs multiple overlapping spatial views of an input image. Each view is independently processed by DINOv3, and overlapping regions are aggregated to enhance feature consistency and boundary robustness.On top of the extracted multi-view features, we further propose a Multi-scale Reconstruction Fusion Transformer (MRFT) to model normal feature distributions via cosine similarity-based reconstruction learning. The anomaly score is computed based on reconstruction residuals, enabling effective discrimination between normal and defective regions.Extensive experiments on three benchmark datasets, including MVTec AD, BTAD, and the Ultrachip dataset, demonstrate that the proposed method consistently outperforms strong baselines in both image-level and pixel-level anomaly detection tasks. The results validate the effectiveness of combining multi-view inference with frozen vision transformer representations for robust industrial anomaly detection.The code will be released at https://github.com/chaoqunWang666.}

%%================================%%
%% Sample for structured abstract %%
%%================================%%

% \abstract{\textbf{Purpose:} The abstract serves both as a general introduction to the topic and as a brief, non-technical summary of the main results and their implications. The abstract must not include subheadings (unless expressly permitted in the journal's Instructions to Authors), equations or citations. As a guide the abstract should not exceed 200 words. Most journals do not set a hard limit however authors are advised to check the author instructions for the journal they are submitting to.
% 
% \textbf{Methods:} The abstract serves both as a general introduction to the topic and as a brief, non-technical summary of the main results and their implications. The abstract must not include subheadings (unless expressly permitted in the journal's Instructions to Authors), equations or citations. As a guide the abstract should not exceed 200 words. Most journals do not set a hard limit however authors are advised to check the author instructions for the journal they are submitting to.
% 
% \textbf{Results:} The abstract serves both as a general introduction to the topic and as a brief, non-technical summary of the main results and their implications. The abstract must not include subheadings (unless expressly permitted in the journal's Instructions to Authors), equations or citations. As a guide the abstract should not exceed 200 words. Most journals do not set a hard limit however authors are advised to check the author instructions for the journal they are submitting to.
% 
% \textbf{Conclusion:} The abstract serves both as a general introduction to the topic and as a brief, non-technical summary of the main results and their implications. The abstract must not include subheadings (unless expressly permitted in the journal's Instructions to Authors), equations or citations. As a guide the abstract should not exceed 200 words. Most journals do not set a hard limit however authors are advised to check the author instructions for the journal they are submitting to.}

\keywords{Industrial Anomaly Detection, Vision Transformers, Feature Reconstruction, Self-supervised Learning}

%%\pacs[JEL Classification]{D8, H51}

%%\pacs[MSC Classification]{35A01, 65L10, 65L12, 65L20, 65L70}

\maketitle

\section{Introduction}\label{sec1}

Industrial Anomaly Detection (IAD) is a fundamental task in automated visual inspection systems, aiming to identify defects from images without requiring large-scale anomaly annotations. In real industrial scenarios, such as ultrasonic chip inspection, defects are typically small-scale, irregular, and embedded in highly noisy grayscale backgrounds. Moreover, spatial misalignment and appearance variability further increase the difficulty of robust defect localization. These challenges require models to simultaneously achieve strong representation capability, spatial robustness, and computational efficiency.

% -----------------------------
% 2. Existing Methods
% -----------------------------
Recent advances in IAD are largely driven by pretrained vision foundation models. Representative paradigms include memory bank-based methods \cite{lendering2026subspacead,lin2026commonality,wang2025search,xu2026mrad}, feature distillation approaches \cite{tien2023revisiting}, and reconstruction-based methods \cite{deng2022anomaly}. Although these methods achieve promising results, they still suffer from several limitations. Memory-based methods introduce high storage and retrieval costs. Distillation-based approaches depend heavily on teacher-student alignment quality. Reconstruction-based frameworks often rely on complex architectural designs such as masked image modeling or heavy encoder-decoder structures, which significantly increase computational overhead and may degrade local structural fidelity.

More recently, transformer-based pretrained models, particularly DINO-style vision transformers, have demonstrated strong potential for industrial anomaly detection due to their powerful semantic representation capability\cite{hou2026visualad}. However, existing approaches typically rely on single-grid patch partitioning during inference, which leads to boundary information loss and insufficient local context modeling. As a result, subtle defects located at patch boundaries are often missed or weakened in the final anomaly response.
\begin{figure*}[!ht]
  \centering
  \includegraphics[width=\textwidth]{secmotivation.png} 
  \caption{Motivation of the proposed framework. \textbf{Left panel:} Visual comparison of anomaly localization on a subtle defect. Conventional single-view inference (bottom) fails to capture the tiny defect due to boundary information loss, whereas our Multi-view Patch Inference (MVPI) strategy (top) achieves accurate localization. \textbf{Right panel:} Comparison of GPU memory usage (GB) across MVTec, Ultrachip, and BTAD datasets. Benefiting from our lightweight Multi-scale Reconstruction Fusion Transformer (MRFT), our method drastically reduces computational overhead compared to DiNomaly and Subspace, ensuring high deployment efficiency.}
  \label{fig:motivation}
\end{figure*}
% -----------------------------
% 3. Motivation
% -----------------------------

To address the above limitations, we revisit the design of patch-based inference and feature reconstruction in a more lightweight and efficient manner. Instead of introducing complex masking strategies or heavy decoder structures, we argue that robust anomaly detection can be achieved through better spatial sampling and hierarchical feature refinement on frozen vision representations.

In this work, we propose a simple yet effective framework built upon a frozen DINOv3 backbone\cite{simeoni2025dinov3}. To enhance spatial robustness, we introduce a Multi-view Patch Inference (MVPI) strategy, which generates multiple overlapping spatial views of an input image. Each view is independently processed by the DINOv3 backbone, and overlapping regions are aggregated to improve boundary consistency and reduce spatial aliasing effects. As illustrated in the left panel of Fig.~\ref{fig:motivation}, by mitigating boundary information loss, our MVPI strategy successfully localizes subtle and hard defects (e.g., the tiny scratch on a screw tip) that are completely missed by conventional single-view methods.

Furthermore, we design a Multi-scale Reconstruction Fusion Transformer (MRFT), which performs hierarchical feature refinement across multiple transformer stages. Specifically, shallow, intermediate, and deep reconstruction layers are used to capture texture-level, structure-level, and semantic-level information, respectively. The multi-scale outputs are then fused to produce a stable reconstruction of normal feature representations. Crucially, unlike existing paradigms that incur heavy computational burdens, our lightweight multi-scale transformer architecture significantly reduces computational overhead. As demonstrated in the right panel of Fig.~\ref{fig:motivation}, our framework drastically lowers GPU memory usage across multiple benchmarks (MVTec, UltraChip, and BTAD) compared to state-of-the-art counterparts, achieving an optimal trade-off between detection accuracy and deployment efficiency.

% -----------------------------
% 4. Contributions
% -----------------------------
The main contributions of this work are summarized as follows:

\begin{itemize}
    \item \textbf{Multi-view Patch Inference Strategy:} We propose a Multi-view Patch Inference (MVPI) mechanism that constructs overlapping spatial views for each input image. This design mitigates boundary information loss in conventional patch-based inference and improves robustness for small and boundary-located defects.

    \item \textbf{Multi-scale Reconstruction Fusion Transformer:} We introduce a lightweight MRFT module that performs hierarchical feature reconstruction across multiple transformer layers. By integrating shallow-to-deep representations, the model achieves more stable and discriminative normal feature modeling.

    \item \textbf{Efficient DINOv3-based Anomaly Detection Framework:} We build a simple yet powerful anomaly detection framework based on a frozen DINOv3 backbone. The proposed method achieves strong performance on MVTec AD, BTAD, and the challenging Ultrachip dataset, demonstrating the effectiveness of combining multi-view inference with multi-scale feature reconstruction.
\end{itemize} 

\section{Related Work}\label{sec2}

\subsection{Large Language and Vision-Language Models for Anomaly Detection}
The emergence of large language models (LLMs) and vision-language models (VLMs) has opened a new paradigm for zero-shot and few-shot industrial anomaly detection. Models such as CLIP\cite{zhou2024anomalyclip}, GPT-4V\cite{gu2024anomalygpt}, and LLaVA leverage massive image-text pretraining to align visual and linguistic representations, enabling anomaly detection through natural language prompts without any task-specific training data. In the zero-shot setting, these models directly compare test images against textual descriptions of "normal" or "anomalous" states via prompt engineering, bypassing the need for normal reference samples entirely\cite{fuvcka2026anomalyvfm}. For few-shot scenarios, in-context learning capabilities allow LLMs to reason about anomalies using only a handful of annotated examples as contextual cues. Recent works further enhance this paradigm by integrating domain-specific knowledge bases, chain-of-thought reasoning, or adapter tuning to bridge the semantic gap between open-world pretraining and fine-grained industrial defect patterns. Despite their remarkable flexibility, VLM-based methods still face challenges in pixel-level localization precision and computational efficiency, often requiring complementary mechanisms to meet real-time industrial deployment requirements.
\subsection{Unsupervised Anomaly Detection}
Unsupervised anomaly detection methods operate under the assumption that only normal samples are available during training. This paradigm is broadly categorized into unified multi-class approaches and single-class approaches, each addressing distinct deployment scenarios.
\subsubsection{Unified Multi-Class Unsupervised Methods}
Unified multi-class methods aim to train a single model capable of detecting anomalies across multiple product categories or domains simultaneously, eliminating the need for category-specific retraining. Representative works leverage shared feature spaces, prompt tuning, or adapter modules to align diverse normal distributions into a common reference frame\cite{guo2025dinomaly,he2024mambaad,lee2026uniformaly,lu2023hierarchical,roming2026superadd,you2022unified},. These approaches significantly reduce deployment overhead in multi-product manufacturing lines but often face trade-offs between cross-category generalization and per-category sensitivity, particularly when inter-class variance is high.
\subsubsection{Single-Class Unsupervised Methods}
Single-class methods remain the dominant paradigm in IAD due to their superior performance on individual product lines. They are further divided into four sub-categories:
\paragraph{Teacher-Student Architectures.}
Teacher-Student (T-S) frameworks employ a pre-trained teacher network to extract target features from normal samples, while a student network is trained to mimic these representations exclusively on normal data \cite{tien2023revisiting}. Anomalies are localized via feature discrepancy between the two networks at inference time. Although T-S models avoid explicit memory banks, they suffer from capacity mismatch challenges: excessive student capacity leads to over-generalization on anomalous regions, diminishing anomaly score distinctiveness, while insufficient capacity hinders faithful feature alignment.
\paragraph{Memory Bank-based Methods.}
These methods construct reference profiles from normal image embeddings to enable non-parametric anomaly scoring. PaDiM \cite{defard2021padim} models multivariate Gaussian distributions of patch features, while PatchCore \cite{roth2022towards} builds a coreset of normal features for nearest-neighbor retrieval. Despite exceptional benchmark performance, memory bank approaches incur substantial GPU memory costs for storing large feature banks and suffer from high inference latency due to pixel-wise neighbor search. Statistical variants also rely on ideal distribution assumptions that may fail in noisy industrial environments.
\paragraph{One-Class Classification (OCC).}
OCC methods learn a compact boundary or hypersphere enclosing normal feature representations in latent space\cite{liu2023simplenet}. Classical approaches like Deep SVDD minimize the volume of a bounding hypersphere around normal embeddings, treating points outside as anomalies. Recent extensions integrate deep feature extractors and adaptive margin mechanisms to handle complex manifolds. However, OCC methods often struggle with high-dimensional feature spaces where the curse of dimensionality causes distance metrics to lose discriminative power, and they typically lack fine-grained localization capability without additional architectural modifications.
\paragraph{Reconstruction-based Methods.}
Reconstruction methods train autoencoders or GANs \cite{fang2025boosting} exclusively on normal samples \cite{iqbal2024multi}, identifying anomalies via input-reconstruction discrepancies. Early approaches suffer from the "identity mapping" problem, where networks generalize too well to anomalous inputs. Advanced methods like DRAEM \cite{zavrtanik2021draem}, CutPaste \cite{li2021cutpaste}, and RealNet \cite{zhang2024realnet} mitigate this through pseudo-anomaly generation or diffusion-synthesized anomalies, but introduce cumbersome augmentation pipelines and computational overhead. In contrast, our approach adheres to a minimalist pure reconstruction philosophy: operating entirely in semantic feature space and explicitly minimizing cosine distance on a unit hypersphere, we bypass identity mapping without relying on complex pseudo-anomaly generation.

\section{Method}\label{sec3}

\subsection{Overview}
We propose a Multi-view Patch Inference framework with a Multi-scale Reconstruction Fusion Transformer (MRFT) for industrial anomaly detection,as illustrated in Fig.\ref{fig:framesec}. The framework consists of three components: (1) a frozen DINOv3 backbone for hierarchical feature extraction, (2) a Multi-view Patch Inference (MVPI) strategy for robust spatial representation, and (3) a multi-scale reconstruction fusion module for anomaly modeling.
\begin{figure*}[!ht]
  \centering
  \includegraphics[width=\textwidth]{secframework.png} 
  \caption{The overall architecture of our Multi-view Patch Inference with a Multi-scale Reconstruction Fusion Transformer Framework}.
  \label{fig:framesec}
\end{figure*}
Given an input image $x$, the model first extracts multi-layer features using a pretrained DINOv3 backbone. Then, multi-view patch inference is applied to enhance spatial robustness. Finally, multi-scale transformer-based reconstruction is performed to estimate normal feature distributions and compute anomaly scores.

\subsection{DINOv3 Foundation Model with Enhanced Token Semantics}
We adopt a pretrained DINOv3 vision transformer as a frozen feature extractor. Compared with earlier DINO versions, DINOv3 introduces several key architectural improvements that are particularly beneficial for industrial anomaly detection.

\subsubsection{Register Tokens for Global Semantic Anchoring}
DINOv3 introduces additional register tokens, which act as global semantic anchors during self-attention computation. Unlike standard patch tokens, register tokens do not correspond to specific spatial locations but instead aggregate global context information. Formally, given input tokens $\mathbf{T} \in \mathbb{R}^{N \times d}$, register tokens $\mathbf{R} \in \mathbb{R}^{k \times d}$ are appended:
\begin{equation}
    \mathbf{T}' = [\mathbf{T}; \mathbf{R}]
\end{equation}
Self-attention is then computed over the extended token set:
\begin{equation}
    \text{Attn}(\mathbf{Q}, \mathbf{K}, \mathbf{V}) = \text{softmax}\left(\frac{\mathbf{Q}\mathbf{K}^\top}{\sqrt{d}}\right)\mathbf{V}
\end{equation}
where register tokens improve global dependency modeling and reduce spatial ambiguity in defect-prone regions.

\subsubsection{High-Capacity Backbone (7B-scale Representation)}
DINOv3 benefits from a significantly larger parameter scale compared to previous versions, reaching up to billions of parameters, enabling stronger representation of fine-grained textures and structural variations. This large-scale pretraining improves texture sensitivity, structural consistency, and cross-category generalization, which are critical for industrial anomaly detection under domain shift.

\subsubsection{Multi-layer Feature Selection Strategy}
We extract intermediate features from layers $l \in \{2, 5, 8\}$:
\begin{itemize}
    \item \textbf{Layer 2 (Low-level features):} Captures edge structures and fine texture gradients; sensitive to micro-defects.
    \item \textbf{Layer 5 (Mid-level features):} Encodes part-level structure and local geometric consistency; handles defect boundary transitions.
    \item \textbf{Layer 8 (High-level features):} Represents semantic consistency and object-level structure; enables global anomaly suppression.
\end{itemize}
The fusion of these layers provides a complementary representation spanning from local texture to global semantics.

\subsection{Multi-view Patch Inference (MVPI)}
To address boundary information loss in conventional patch-based inference, we propose a Multi-view Patch Inference strategy implemented at the feature extraction stage,as illustrated in Fig.\ref{fig:framesec} (a).

\subsubsection{Spatial Multi-view Construction}
Given an input feature map $\mathbf{X}$, we construct a set of overlapping views $\mathcal{V} = \{x_1, x_2, \dots, x_N\}$, where each view is generated via a sliding-window mechanism with overlap constraint:
\begin{equation}
    x_i = \mathcal{C}(\mathbf{X}; s_i, s_i + p)
\end{equation}
Here, $p$ denotes the patch size, $s_i$ represents the spatial offset with overlap, and the overlap ratio controls the redundancy level.

\subsubsection{Multi-view Feature Encoding}
Each view is independently processed by the frozen DINOv3 backbone:
\begin{equation}
    \mathbf{F}_i^l = \phi_l(x_i), \quad l \in \{2, 5, 8\}
\end{equation}
This results in redundant feature observations of the same spatial region under different receptive contexts.

\subsubsection{Cross-view Feature Aggregation}
For overlapping regions, features are aggregated using weighted averaging:
\begin{equation}
    \mathbf{F} = \sum_{i=1}^{N} w_i \mathbf{F}_i
\end{equation}
where weights $w_i$ are implicitly determined by spatial coverage frequency. This mechanism improves boundary consistency, robustness to spatial misalignment, and sensitivity to small defects.

\subsubsection{Effect of MVPI}
MVPI effectively transforms single-view inference into an implicit ensemble of multiple spatial views, reducing the risk of missing anomalies located at patch boundaries.

\subsection{Multi-scale Reconstruction Fusion Transformer (MRFT)}

\subsubsection{Motivation}
Single-scale reconstruction methods are limited in capturing hierarchical feature dependencies. To address this, we propose a multi-scale transformer-based reconstruction strategy,as illustrated in Fig.\ref{fig:framesec} (b).

\subsubsection{Hierarchical Reconstruction Formulation}
Given input token features $\mathbf{z}$, we define a three-stage transformation:
\begin{equation}
    \mathbf{z}_1 = f_1(\mathbf{z}), \quad \mathbf{z}_2 = f_2(\mathbf{z}_1), \quad \mathbf{z}_3 = f_3(\mathbf{z}_2)
\end{equation}
where each $f_i$ is a lightweight self-attention encoder layer:
\begin{equation}
    f_i(\mathbf{x}) = \text{LN}(\mathbf{x} + \text{MSA}(\mathbf{x})) + \text{FFN}(\mathbf{x})
\end{equation}
where LN denotes Layer Normalization, MSA denotes Multi-head Self-Attention, and FFN denotes Feed-Forward Network.

\subsubsection{Multi-scale Feature Fusion}
The final reconstruction output is obtained via fusion:
\begin{equation}
    \hat{\mathbf{z}} = \sum_{i=1}^{3} \alpha_i \mathbf{z}_i, \quad \sum_{i=1}^{3} \alpha_i = 1
\end{equation}
This design enables low-level refinement (texture), mid-level correction (structure), and high-level stabilization (semantics).

\subsubsection{Jitter Regularization}
To improve robustness, we introduce feature jitter:
\begin{equation}
    \mathbf{z}' = \mathbf{z} + \epsilon, \quad \epsilon \sim \mathcal{N}(0, \sigma^2)
\end{equation}
which prevents overfitting to exact normal feature patterns.

\subsection{Anomaly Score Computation}
The anomaly score is defined as the reconstruction deviation:
\begin{equation}
    S = 1 - \cos(\mathbf{z}, \hat{\mathbf{z}})
\end{equation}
Pixel-level anomaly maps are obtained by reshaping token-wise scores and interpolating to image resolution. The image-level anomaly score is computed via top-$k$ aggregation:
\begin{equation}
    S_{\text{img}} = \frac{1}{K} \sum_{i \in \text{TopK}} S_i
\end{equation}
Gaussian smoothing is subsequently applied to improve spatial continuity.

\section{Experiments}\label{sec4}
\subsection{Experimental Setup}

\textbf{Datasets.} 
To comprehensively evaluate the robustness and generalizability of our proposed method, we conduct extensive experiments on three distinct industrial anomaly detection datasets. 
1) \textbf{MVTec AD}\cite{bergmann2019mvtec}: A standard benchmark containing 15 categories of industrial products, widely used to evaluate basic detection performance. 
2) \textbf{BTAD}\cite{mishra2021vt}: Comprises 3 categories of real-world industrial scenarios, heavily utilized to assess the model's environmental robustness. 
3) \textbf{UltraChip}\cite{11372587}: A highly challenging dataset focusing on micro-defect detection in semiconductor chips, which we employ to validate our method's generalization capability on extremely fine-grained microscopic anomalies.

\textbf{Evaluation Metrics.}
Following standard anomaly detection protocols, we employ three widely-accepted metrics to quantitatively evaluate our model. We use Image-level Area Under the Receiver Operating Characteristic Curve (I-AUC) for anomaly classification, Pixel-level Area Under the ROC (P-AUC) to assess anomaly localization precision, and Per-Region Overlap (PRO) to measure the region-wise overlap rate of the predicted anomalous areas.

\textbf{Implementation Details.}
Most of our experiments are conducted on a single NVIDIA V100 GPU (24GB).All input images are uniformly resized to $448 \times 448$. To construct a multi-scale receptive field, we extract intermediate features from the 2nd, 5th, and 8th layers of the encoder.During the inference stage, we perform a topological smoothing operation using a Gaussian filter with a constant standard deviation ($\sigma=1.5$) to suppress background noise, and establish robust image-level anomaly scores by calculating the mean of the Top-$K$ ($K=100$) extreme values.

\subsection{Comparison with State-of-the-Art Methods}
To comprehensively evaluate our framework, we integrate quantitative metrics and qualitative visualizations. We first benchmark our method on the highly challenging real-world dataset, and subsequently verify its generalizability on standard IAD benchmarks.

\textbf{Performance on Real-World UltraChip Dataset.}
We prioritize the evaluation on the UltraChip dataset, which poses extreme challenges due to severe structural noise, limited labeled data, and highly diverse, visually imperceptible microscopic defects. As summarized in Table \ref{tab:ultrachip_detail}, while traditional methods suffer sharp performance degradation from heavy background interference, our method successfully maintains a state-of-the-art mean Image-AUC of 81.9\%. Notably, our approach outperforms recent models like URD and SuperSimpleNet by significant margins. Furthermore, as demonstrated perfectly in the qualitative results in Figure \ref{fig:ultrachip_vis}, our framework precisely highlights actual defect areas without being inundated by widespread false-positive background noise (avoiding massive "red" regions), proving its exceptional capability to extract intrinsic normal patterns in non-ideal industrial environments.
\begin{table}[!ht]
\centering
\caption{Detailed quantitative comparison on the UltraChip dataset. We report the performance across all three specific chip categories.}
\label{tab:ultrachip_detail}
\small % 1. 缩小字号，这是在不改动表头的情况下最有效的办法
\setlength{\tabcolsep}{3.5pt} % 2. 减小列间距，默认通常是 6pt，这里压缩到 3.5pt
\begin{tabular}{@{}lcccccccccccc@{}}
\toprule
Class & \multicolumn{3}{c}{Four-Lead} & \multicolumn{3}{c}{Bonding} & \multicolumn{3}{c}{Two-Lead} & \multicolumn{3}{c}{Mean} \\
\cmidrule(lr){2-4} \cmidrule(lr){5-7} \cmidrule(lr){8-10} \cmidrule(l){11-13}
Method & I-AUC & P-AUC & PRO & I-AUC & P-AUC & PRO & I-AUC & P-AUC & PRO & I-AUC & P-AUC & PRO \\
\midrule
REB\cite{lyu2024reb} & 64.0 & 83.2 & 33.9 & 83.2 & 0.285 & 43.6 & 66.1 & 87.3 & 33.2 & 71.1 & 68.3 & 36.9 \\
FUNAD\cite{im2025fun} & 72.5 & 82.5 & 39.0 & 85.1 & 94.8 & 44.5 & 76.4 & 81.4 & 29.5 & 78.0 & 86.2 & 37.6 \\
RAS\cite{yang2024context} & 77.9 & 87.7 & 36.0 & 64.9 & 80.6 & 44.4 & 77.7 & 76.4 & 29.2 & 73.5 & 81.6 & 36.5 \\
MSFR\cite{iqbal2024multi} & 36.7 & 55.4 & 19.6 & 53.6 & 57.4 & 16.4 & 50.1 & 51.5 & 18.5 & 46.8 & 54.8 & 18.1 \\
FOD\cite{yao2023focus} & 79.1 & 80.2 & 32.6 & 79.9 & 89.0 & 54.7 & 79.8 & 74.9 & 18.6 & 79.6 & 81.3 & 35.3 \\
RD++\cite{tien2023revisiting} & 71.9 & 84.7 & 50.8 & 64.2 & 86.1 & 41.4 & 77.7 & 83.9 & 44.7 & 71.2 & 84.9 & 45.6 \\
HGAD\cite{yao2024hierarchical} & 74.3 & 83.5 & 46.6 & 78.1 & 88.8 & 43.1 & 75.8 & 83.2 & 27.5 & 76.0 & 85.1 & 39.1 \\
SuperSimpleNet\cite{rolih2024supersimplenet} & 72.9 & 67.1 & 40.3 & 78.0 & 81.5 & 56.2 & 79.4 & 68.7 & 33.2 & 76.8 & 72.5 & 43.2 \\
URD\cite{liu2025unlocking} & 66.7 & 82.6 & 29.4 & 86.8 & 95.5 & 62.8 & 68.8 & 79.2 & 21.5 & 74.1 & 85.7 & 37.9 \\
DiNomaly\cite{guo2025dinomaly} & 83.9 & 83.9 & 47.6 & 64.9 & 87.6 & 48.1 & 80.3 & 84.6 & 44.4 & 76.4 & 85.4 & 46.7 \\
Subspace\cite{lendering2026subspacead} & 73.5 & 85.5 & 34.1 & 27.6 & 73.1 & 42.9 & 55.7 & 85.3 & 29.6 & 52.3 & 81.3 & 35.5 \\
\textbf{Ours} & \textbf{80.1} & \textbf{84.3} & \textbf{45.7} & \textbf{82.5} & \textbf{92.3} & \textbf{58.9} & \textbf{83.3} & \textbf{82.2} & \textbf{34.6} & \textbf{81.9} & \textbf{86.3} & \textbf{46.4} \\
\botrule
\end{tabular}
\end{table}

\begin{figure*}[!htbp]
  \centering
  \includegraphics[width=0.95\textwidth, keepaspectratio]{fig6.png}
  \caption{Qualitative comparison of anomaly localization on the challenging UltraChip dataset.}
  \label{fig:ultrachip_vis}
\end{figure*}
\textbf{Generalizability on Standard Benchmarks (MVTec AD \& BTAD).}
To prove that our method does not overfit to specific noisy datasets, we benchmark it on the standard MVTec AD and BTAD datasets. As presented in Table \ref{tab:mvtec_results} \ref{tab:auc_comparison} , despite its architectural minimalism, our method achieves the most comprehensive performance on MVTec AD, securing the highest scores across all two metrics (99.2\% P-AUC, and 96.0\% PRO). This represents a clear improvement over SOTA methods like Dinomaly and Subspace. To intuitively demonstrate this, Figure \ref{fig:mvtec_vis} shows that our generated heatmaps align tightly with ground-truth masks with almost zero background noise. Moreover, Table \ref{tab:btad_results} confirms our robustness against varying lighting and complex topologies typical of factory settings in the BTAD dataset, where we maintain a consistent advantage in average metrics (97.3\% I-AUC).
\begin{table}[!htbp]
\centering
\caption{Quantitative comparison on the MVTec AD dataset.}
\label{tab:mvtec_results}
% 此处 0.85\columnwidth 控制表格大小，可自行微调（如 0.9 或 0.8）
\begin{tabular}{lccc}
\toprule
Method & I-AUC & P-AUC & PRO \\
\midrule
REB\cite{lyu2024reb} & 99.1 & 98.4 & 92.5 \\
URD\cite{liu2025unlocking} & 98.7 & 98.8 & 96.0 \\
SuperSimpleNet\cite{rolih2024supersimplenet} & 97.4 & 98.5 & 91.0 \\
FUNAD\cite{im2025fun} & 99.2 & 98.1 & 94.2 \\
RD++\cite{tien2023revisiting} & 98.2 & 99.2 & 94.9 \\
FOD\cite{yao2023focus} & 98.9 & 98.3 & 93.5 \\
HGAD\cite{yao2024hierarchical} & 98.4 & 97.9 & 94.2 \\
RAS\cite{yang2024context} & 98.5 & 97.5 & 93.9 \\
MSFR\cite{iqbal2024multi} & 98.4 & 97.3 & 93.0 \\
Dinomaly \cite{guo2025dinomaly} & 99.6 & 98.3 & 94.8 \\
Subspace \cite{lendering2026subspacead} & 99.1 & 98.1 & 94.9 \\
\midrule
\textbf{Ours} & \textbf{99.5} & \textbf{99.2} & \textbf{96.0} \\
\bottomrule
\end{tabular}
\end{table}


% 新增的 MVTec 可视化图片代码 (如果图片较大，可将 figure 换为 \begin{figure*}[htpb])
\begin{figure}[!htbp]
  \centering
  \includegraphics[width=\linewidth]{fig4.png} % 请将此处替换为您的实际图片名称和路径
  \caption{Qualitative anomaly localization results of our proposed method on the MVTec AD dataset. For each category, we present the raw input image, the ground-truth anomaly mask, and the highly precise anomaly heatmap generated by our framework.}
  \label{fig:mvtec_vis}
\end{figure}
\begin{table}[htbp]
    \centering
    \caption{MVTec Performance Metrics (AUC) --- DINOv3 vs. DINOv3+MVPI Results}
    \label{tab:auc_comparison}
    \renewcommand{\arraystretch}{1.2}
    \begin{tabular}{lcccccc}
        \toprule
        \textbf{Category} & 
        \multicolumn{3}{c}{\textbf{DINOv3}} & 
        \multicolumn{3}{c}{\textbf{DINOv3+MVPI}} \\
        \cmidrule(lr){2-4} \cmidrule[\heavyrulewidth](lr){5-7}
        & \textbf{I-AUC} & \textbf{P-AUC} & \textbf{PRO} &
          \textbf{I-AUC} & \textbf{P-AUC} & \textbf{PRO} \\
        \midrule
        hazelnut   & 100.0 & 99.5 & 96.7 & 100.0 & 99.5 & 96.9 \\
        metal\_nut & 100.0 & 99.4 & 97.1 & 100.0 & 99.3 & 97.0 \\
        bottle     & 100.0 & 99.2 & 94.0 & 100.0 & 99.2 & 97.0 \\
        cable      & 99.7  & 99.2 & 99.0 & 99.7  & 99.1 & 99.5 \\
        capsule    & 98.4  & 99.3 & 89.0 & 98.7  & 99.3 & 89.2 \\
        carpet     & 100.0 & 99.7 & 98.3 & 100.0 & 99.7 & 98.0 \\
        grid       & 100.0 & 99.1 & 96.3 & 100.0 & 99.2 & 96.0 \\
        leather    & 100.0 & 99.4 & 98.4 & 100.0 & 99.5 & 98.5 \\
        pill       & 98.3  & 98.7 & 96.9 & 98.6  & 99.0 & 97.1 \\
        screw      & 95.1  & 99.3 & 95.8 & 98.6  & 99.5 & 96.5 \\
        tile       & 100.0 & 98.8 & 95.5 & 100.0 & 98.8 & 95.7 \\
        toothbrush & 100.0 & 99.2 & 89.8 & 100.0 & 99.3 & 92.5 \\
        transistor & 99.7  & 98.7 & 96.3 & 99.7  & 98.5 & 96.3 \\
        wood       & 97.5  & 97.0 & 92.1 & 97.5  & 97.0 & 92.3 \\
        zipper     & 99.9  & 98.9 & 96.4 & 99.9  & 99.3 & 97.6 \\
        \midrule[\heavyrulewidth]
        \textbf{Mean} & \textbf{99.2} & \textbf{99.0} & \textbf{95.4} &
                        \textbf{99.5} & \textbf{99.2} & \textbf{96.0} \\
        \bottomrule
    \end{tabular}
\end{table}

% Table 2: BTAD Results
\begin{table}[htbp] 
\centering
\caption{Quantitative comparison on the BTAD dataset per class.}
\label{tab:btad_results}
\begin{tabular}{c|ccc|ccc|ccc|ccc}
\hline
Class & \multicolumn{3}{c|}{Class 01} & \multicolumn{3}{c|}{Class 02} & \multicolumn{3}{c|}{Class 03} & \multicolumn{3}{c}{Average} \\
\hline
Method & I-AUC & P-AUC & PRO & I-AUC & P-AUC & PRO & I-AUC & P-AUC & PRO & I-AUC & P-AUC & PRO \\
\hline
FOD\cite{yao2023focus} & 95.8 & 97.0 & 62.2 & 96.2 & 97.8 & 65.4 & 96.0 & 97.7 & 99.0 & 96.0 & 97.5 & 75.5 \\
Subspace\cite{lendering2026subspacead} & 100 & 98.9 & 84.7 & 91.2 & 96.8 & 80.4 & 99.6 & 98.5 & 56.1 & 96.9 & 98.0 & 73.7 \\
Dinomaly\cite{guo2025dinomaly} & 96.4 & 97.2 & 74.0 & 89.1 & 96.5 & 56.6 & 99.8 & 99.8 & 99.3 & 95.1 & 97.9 & 76.6 \\
REB\cite{lyu2024reb} & 99.6 & 94.7 & 70.0 & 88.5 & 95.6 & 66.4 & 99.8 & 99.7 & 80.6 & 96.0 & 97.2 & 72.4 \\
HGAD\cite{yao2024hierarchical} & 92.6 & 94.9 & 70.5 & 93.3 & 96.7 & 59.3 & 97.0 & 99.1 & 97.9 & 94.1 & 96.9 & 75.9 \\
RAS\cite{yang2024context} & 92.7 & 96.9 & 70.3 & 92.3 & 94.7 & 55.5 & 99.1 & 99.0 & 99.0 & 94.7 & 97.0 & 74.9 \\
\textbf{Ours} & \textbf{96.5} & \textbf{96.8} & \textbf{63.6} & \textbf{95.5} & \textbf{97.8} & \textbf{67.5} & \textbf{99.9} & \textbf{99.9} & \textbf{99.0} & \textbf{97.3} & \textbf{98.2} & \textbf{76.7} \\
\hline
\end{tabular}
\end{table}
\subsection{Ablation Studies}
To rigorously validate the efficacy of each proposed module and verify the rationality of our structural design, we conduct comprehensive ablation studies on the MVTec AD datasets.
\begin{table}[!ht]
    \centering
    \caption{Ablation study on feature layer selection for the \textit{screw} category.}
    \label{tab:ablation_screw}
    \renewcommand{\arraystretch}{1.2} % 增加行高，使表格更易读
    \begin{tabular}{cccc}
        \toprule
        \textbf{Feature Layers} & \textbf{I-AUC} & \textbf{P-AUC} & \textbf{PRO} \\
        \midrule
        $\{3, 6, 9\}$           & 93.9          & \textbf{99.3} & 95.6 \\
        $\{4, 7, 10\}$          & 92.7          & 99.2          & 95.7 \\
        $\{2, 5, 8\}$           & \textbf{95.2} & 99.2          & \textbf{95.8} \\
        \bottomrule
    \end{tabular}
\end{table}

% ==================== 分析段落部分 ====================
\noindent
As demonstrated in Table~\ref{tab:ablation_screw}, we conducted an ablation study to evaluate the impact of different feature layer combinations from DINOv3 on anomaly detection performance for the \textit{screw} category. 
The configuration utilizing layers $\{2, 5, 8\}$ achieves the highest I-AUC (95.2\%) and PRO (95.8\%), while maintaining a competitive P-AUC (99.2\%). 
Although the $\{3, 6, 9\}$ setting yields a marginally superior P-AUC (99.3\%), it suffers from a noticeable degradation in image-level discrimination (I-AUC drops by 1.3\%). 
We attribute the superior overall performance of layers $\{2, 5, 8\}$ to their balanced representation: this specific combination effectively captures fine-grained shallow texture and structural details from the early layers, while simultaneously preserving rich high-level semantic information from the deeper layers. 
Consequently, we adopt $\{2, 5, 8\}$ as the optimal feature layer configuration for DINOv3 in our final model.





\textbf{Rationale for Multi-Scale Layer Selection.}
Given the diverse manifestations of defects—from tiny cracks to large structural missing parts—single-scale features are highly inadequate. We meticulously select features from the 2nd, 5th, and 8th layers of the frozen DINOv3 backbone. As demonstrated in Table \ref{tab:ablation_layers}, this $\{2, 5, 8\}$ configuration yields the best results (99.5\% I-AUC unconditionally). The underlying rationale is complementary: Layer 2 is shallow enough to capture fine-grained textural anomalies and high-frequency noise typical in UltraChip; Layer 5 acts as an intermediary bridge for semantic transitions; Layer 8 is deep enough to capture local structural semantics without becoming too abstract or losing spatial boundary resolution. In contrast, using only deeper layers like $\{5, 8\}$ drops the average I-AUC to 98.30\% due to the loss of fine-grained details.concepts, ensuring holistic anomaly localization.

% Table : Multi-scale layer selection
\begin{table}[!ht]
\centering
\caption{Performance comparison of multi-scale layer selections on the MVTec AD dataset.}
\label{tab:ablation_layers}
\begin{tabular}{c|c|c|c|c}
\toprule
Selected Layers & Category & I-AUC & P-AUC & PRO \\
\midrule
\multirow{2}{*}{\{5, 8\}} & Average & 98.30 & 97.60 & 95.00 \\
                          & Cable & 94.79 & 98.78 & 97.40 \\
\midrule
\multirow{2}{*}{\textbf{\{2, 5, 8\}}} & Average & \textbf{99.20} & \textbf{99.60} & \textbf{97.40} \\
                                      & Cable & \textbf{99.50} & \textbf{99.30} & \textbf{99.00} \\
\bottomrule
\end{tabular}
\end{table}





% 
\textbf{Superiority of Top-$K$ Score Aggregation.}
Finally, we evaluate strategies for mapping pixel-level anomaly maps to global image-level scores. In real-world industrial scenarios, background reflections and sensor artifacts often induce isolated high-frequency noise spikes. As demonstrated in Table \ref{tab:ablation_aggregation}, naive Global Max-pooling is highly vulnerable to these single-pixel outliers, severely inflating the false positive rate. Conversely, Global Average-pooling over-smooths the score map, drowning out subtle microscopic defects within dominant normal regions. 

It is worth noting that before aggregation, our pipeline applies Gaussian smoothing ($\sigma=1.5$) and squared enhancement to suppress high-frequency jitter and amplify salient defect responses. Building on this refined score map, our proposed Top-$K$ Pooling ($K=100$) leverages a rank-based selection mechanism that inherently excludes isolated noise spikes while preserving the spatial density signals from actual defect regions. This robust aggregation strategy achieves the best trade-off between noise suppression and defect sensitivity, maximizing overall detection performance.

% Table: Score aggregation strategies
\begin{table}[!ht]
    \centering
    \caption{Comparison of image-level score aggregation strategies on MVTec AD. All methods share identical pixel-level preprocessing (Gaussian blur + squared enhancement) before aggregation.}
    \label{tab:ablation_aggregation}
    \renewcommand{\arraystretch}{1.2}
    \begin{tabular}{lccc}
        \toprule
        \textbf{Aggregation Strategy} & \textbf{I-AUC} & \textbf{P-AUC} & \textbf{PRO} \\
        \midrule
        Global Max-pooling            & 98.9          & 98.8          & 96.2        \\
        Global Average-pooling        & 98.3          & 99.2          & 95.0        \\
        \textbf{Top-$K$ Pooling ($K=100$)} & \textbf{99.5} & \textbf{99.1} & \textbf{95.8} \\
        \bottomrule
    \end{tabular}
\end{table}

\section{Conclusion}\label{sec5}
In this paper, we propose a simple yet effective Multi-view Patch Inference with Reconstruction Fusion Transformer (MVPI-RFT) framework for industrial anomaly detection. By introducing a Multi-view Patch Inference (MVPI) strategy and a Multi-scale Reconstruction Fusion Transformer (MRFT) mechanism, our approach effectively addresses the limitations of conventional single-grid patch partitioning while completely eliminating the need for cumbersome memory banks and complex distillation-based architectures. Extensive experiments on standard benchmarks, including MVTec AD, BTAD, and the highly challenging UltraChip dataset, demonstrate that our method achieves state-of-the-art performance in both image-level and pixel-level anomaly detection, with strong robustness and precise localization ability, while maintaining a lightweight and memory-efficient inference pipeline.

Despite these advantages, the use of a frozen DINOv3 backbone introduces additional computational overhead, leading to relatively higher inference latency. Future work will focus on designing more efficient backbone-adaptive strategies to reduce computational cost and improve real-time applicability.
Interestingly, our method also exhibits promising zero-shot generalization behavior on the UltraChip dataset, suggesting that multi-view feature aggregation with frozen vision transformers may implicitly capture transferable normality priors. This observation opens up an interesting direction for extending the framework toward zero-shot or few-shot anomaly detection scenarios.

Lastly, our current framework still operates under a conventional single-class training paradigm, which requires training separate models for different product categories. In future work, we aim to overcome this limitation by developing a unified multi-class anomaly detection framework. We further plan to explore more compact and decoupled feature representation strategies to enhance generalization across diverse industrial categories, while also improving inference efficiency for deployment on edge devices.
\section*{Declarations}

\subsection*{Availability of data and material}
The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.
% 如果数据已公开，请替换为：The dataset is available at [Repository Name] under DOI: [xxx].

\subsection*{Competing interests}
The authors declare that they have no competing interests.
% 如果有利益冲突，请如实声明，例如：Author A is a consultant for Company B.

\subsection*{Funding}
The authors did not receive support from any organization for the submitted work.
% 如果没有资助，请写：The authors did not receive support from any organization for the submitted work.

\subsection*{Authors' contributions}
Chaoqun Wang: Conceptualization, Methodology, Formal Analysis, Writing -- Original Draft, Project Administration. 
Wenjing Zhang: Validation, Investigation, Data Curation, Writing -- Review \& Editing. 
Bo Qi: Resources, Visualization. 
Xiaoyu Huang: Resources, Visualization. 
Ning He: Supervision, Funding Acquisition, Final Approval. 
All authors read and approved the final manuscript.
\subsection{Acknowledgments}
This work was supported by the Key Research and Development Project of the Ministry of Science and Technology: Ultrasonic Microscope (2023YFF0716500) 

% \noindent
% The input format for the above table is as follows:

% %%=============================================%%
% %% For presentation purpose, we have included  %%
% %% \bigskip command. Please ignore this.       %%
% %%=============================================%%
% \bigskip
% \begin{verbatim}
% \begin{table}[<placement-specifier>]
% \caption{<table-caption>}\label{<table-label>}%
% \begin{tabular}{@{}llll@{}}
% \toprule
% Column 1 & Column 2 & Column 3 & Column 4\\
% \midrule
% row 1 & data 1 & data 2	 & data 3 \\
% row 2 & data 4 & data 5\footnotemark[1] & data 6 \\
% row 3 & data 7 & data 8	 & data 9\footnotemark[2]\\
% \botrule
% \end{tabular}
% \footnotetext{Source: This is an example of table footnote. 
% This is an example of table footnote.}
% \footnotetext[1]{Example for a first table footnote.
% This is an example of table footnote.}
% \footnotetext[2]{Example for a second table footnote. 
% This is an example of table footnote.}
% \end{table}
% \end{verbatim}
% \bigskip
% %%=============================================%%
% %% For presentation purpose, we have included  %%
% %% \bigskip command. Please ignore this.       %%
% %%=============================================%%

% \begin{table}[h]
% \caption{Example of a lengthy table which is set to full textwidth}\label{tab2}
% \begin{tabular*}{\textwidth}{@{\extracolsep\fill}lcccccc}
% \toprule%
% & \multicolumn{3}{@{}c@{}}{Element 1\footnotemark[1]} & \multicolumn{3}{@{}c@{}}{Element 2\footnotemark[2]} \\\cmidrule{2-4}\cmidrule{5-7}%
% Project & Energy & $\sigma_{calc}$ & $\sigma_{expt}$ & Energy & $\sigma_{calc}$ & $\sigma_{expt}$ \\
% \midrule
% Element 3  & 990 A & 1168 & $1547\pm12$ & 780 A & 1166 & $1239\pm100$\\
% Element 4  & 500 A & 961  & $922\pm10$  & 900 A & 1268 & $1092\pm40$\\
% \botrule
% \end{tabular*}
% \footnotetext{Note: This is an example of table footnote. This is an example of table footnote this is an example of table footnote this is an example of~table footnote this is an example of table footnote.}
% \footnotetext[1]{Example for a first table footnote.}
% \footnotetext[2]{Example for a second table footnote.}
% \end{table}

% In case of double column layout, tables which do not fit in single column width should be set to full text width. For this, you need to use \verb+\begin{table*}+ \verb+...+ \verb+\end{table*}+ instead of \verb+\begin{table}+ \verb+...+ \verb+\end{table}+ environment. Lengthy tables which do not fit in textwidth should be set as rotated table. For this, you need to use \verb+\begin{sidewaystable}+ \verb+...+ \verb+\end{sidewaystable}+ instead of \verb+\begin{table*}+ \verb+...+ \verb+\end{table*}+ environment. This environment puts tables rotated to single column width. For tables rotated to double column width, use \verb+\begin{sidewaystable*}+ \verb+...+ \verb+\end{sidewaystable*}+.

% \begin{sidewaystable}
% \caption{Tables which are too long to fit, should be written using the ``sidewaystable'' environment as shown here}\label{tab3}
% \begin{tabular*}{\textheight}{@{\extracolsep\fill}lcccccc}
% \toprule%
% & \multicolumn{3}{@{}c@{}}{Element 1\footnotemark[1]}& \multicolumn{3}{@{}c@{}}{Element\footnotemark[2]} \\\cmidrule{2-4}\cmidrule{5-7}%
% Projectile & Energy	& $\sigma_{calc}$ & $\sigma_{expt}$ & Energy & $\sigma_{calc}$ & $\sigma_{expt}$ \\
% \midrule
% Element 3 & 990 A & 1168 & $1547\pm12$ & 780 A & 1166 & $1239\pm100$ \\
% Element 4 & 500 A & 961  & $922\pm10$  & 900 A & 1268 & $1092\pm40$ \\
% Element 5 & 990 A & 1168 & $1547\pm12$ & 780 A & 1166 & $1239\pm100$ \\
% Element 6 & 500 A & 961  & $922\pm10$  & 900 A & 1268 & $1092\pm40$ \\
% \botrule
% \end{tabular*}
% \footnotetext{Note: This is an example of table footnote this is an example of table footnote this is an example of table footnote this is an example of~table footnote this is an example of table footnote.}
% \footnotetext[1]{This is an example of table footnote.}
% \end{sidewaystable}

% \section{Figures}\label{sec6}

% As per the \LaTeX\ standards you need to use eps images for \LaTeX\ compilation and \verb+pdf/jpg/png+ images for \verb+PDFLaTeX+ compilation. This is one of the major difference between \LaTeX\ and \verb+PDFLaTeX+. Each image should be from a single input .eps/vector image file. Avoid using subfigures. The command for inserting images for \LaTeX\ and \verb+PDFLaTeX+ can be generalized. The package used to insert images in \verb+LaTeX/PDFLaTeX+ is the graphicx package. Figures can be inserted via the normal figure environment as shown in the below example:

% %%=============================================%%
% %% For presentation purpose, we have included  %%
% %% \bigskip command. Please ignore this.       %%
% %%=============================================%%
% \bigskip
% \begin{verbatim}
% \begin{figure}[<placement-specifier>]
% \centering
% \includegraphics{<eps-file>}
% \caption{<figure-caption>}\label{<figure-label>}
% \end{figure}
% \end{verbatim}
% \bigskip
% %%=============================================%%
% %% For presentation purpose, we have included  %%
% %% \bigskip command. Please ignore this.       %%
% %%=============================================%%

% \begin{figure}[h]
% \centering
% \includegraphics[width=0.9\textwidth]{fig.eps}
% \caption{This is a widefig. This is an example of long caption this is an example of long caption  this is an example of long caption this is an example of long caption}\label{fig1}
% \end{figure}

% In case of double column layout, the above format puts figure captions/images to single column width. To get spanned images, we need to provide \verb+\begin{figure*}+ \verb+...+ \verb+\end{figure*}+.

% For sample purpose, we have included the width of images in the optional argument of \verb+\includegraphics+ tag. Please ignore this. 

% \section{Algorithms, Program codes and Listings}\label{sec7}

% Packages \verb+algorithm+, \verb+algorithmicx+ and \verb+algpseudocode+ are used for setting algorithms in \LaTeX\ using the format:

% %%=============================================%%
% %% For presentation purpose, we have included  %%
% %% \bigskip command. Please ignore this.       %%
% %%=============================================%%
% \bigskip
% \begin{verbatim}
% \begin{algorithm}
% \caption{<alg-caption>}\label{<alg-label>}
% \begin{algorithmic}[1]
% . . .
% \end{algorithmic}
% \end{algorithm}
% \end{verbatim}
% \bigskip
% %%=============================================%%
% %% For presentation purpose, we have included  %%
% %% \bigskip command. Please ignore this.       %%
% %%=============================================%%

% You may refer above listed package documentations for more details before setting \verb+algorithm+ environment. For program codes, the ``verbatim'' package is required and the command to be used is \verb+\begin{verbatim}+ \verb+...+ \verb+\end{verbatim}+. 

% Similarly, for \verb+listings+, use the \verb+listings+ package. \verb+\begin{lstlisting}+ \verb+...+ \verb+\end{lstlisting}+ is used to set environments similar to \verb+verbatim+ environment. Refer to the \verb+lstlisting+ package documentation for more details.

% A fast exponentiation procedure:

% \lstset{texcl=true,basicstyle=\small\sf,commentstyle=\small\rm,mathescape=true,escapeinside={(*}{*)}}
% \begin{lstlisting}
% begin
%   for $i:=1$ to $10$ step $1$ do
%       expt($2,i$);  
%       newline() od                (*\textrm{Comments will be set flush to the right margin}*)
% where
% proc expt($x,n$) $\equiv$
%   $z:=1$;
%   do if $n=0$ then exit fi;
%      do if odd($n$) then exit fi;                 
%         comment: (*\textrm{This is a comment statement;}*)
%         $n:=n/2$; $x:=x*x$ od;
%      { $n>0$ };
%      $n:=n-1$; $z:=z*x$ od;
%   print($z$). 
% end
% \end{lstlisting}

% \begin{algorithm}
% \caption{Calculate $y = x^n$}\label{algo1}
% \begin{algorithmic}[1]
% \Require $n \geq 0 \vee x \neq 0$
% \Ensure $y = x^n$ 
% \State $y \Leftarrow 1$
% \If{$n < 0$}\label{algln2}
%         \State $X \Leftarrow 1 / x$
%         \State $N \Leftarrow -n$
% \Else
%         \State $X \Leftarrow x$
%         \State $N \Leftarrow n$
% \EndIf
% \While{$N \neq 0$}
%         \If{$N$ is even}
%             \State $X \Leftarrow X \times X$
%             \State $N \Leftarrow N / 2$
%         \Else[$N$ is odd]
%             \State $y \Leftarrow y \times X$
%             \State $N \Leftarrow N - 1$
%         \EndIf
% \EndWhile
% \end{algorithmic}
% \end{algorithm}

% %%=============================================%%
% %% For presentation purpose, we have included  %%
% %% \bigskip command. Please ignore this.       %%
% %%=============================================%%
% \bigskip
% \begin{minipage}{\hsize}%
% \lstset{frame=single,framexleftmargin=-1pt,framexrightmargin=-17pt,framesep=12pt,linewidth=0.98\textwidth,language=pascal}% Set your language (you can change the language for each code-block optionally)
% %%% Start your code-block
% \begin{lstlisting}
% for i:=maxint to 0 do
% begin
% { do nothing }
% end;
% Write('Case insensitive ');
% Write('Pascal keywords.');
% \end{lstlisting}
% \end{minipage}

% \section{Cross referencing}\label{sec8}

% Environments such as figure, table, equation and align can have a label
% declared via the \verb+\label{#label}+ command. For figures and table
% environments use the \verb+\label{}+ command inside or just
% below the \verb+\caption{}+ command. You can then use the
% \verb+\ref{#label}+ command to cross-reference them. As an example, consider
% the label declared for Figure~\ref{fig1} which is
% \verb+\label{fig1}+. To cross-reference it, use the command 
% \verb+Figure \ref{fig1}+, for which it comes up as
% ``Figure~\ref{fig1}''. 

% To reference line numbers in an algorithm, consider the label declared for the line number 2 of Algorithm~\ref{algo1} is \verb+\label{algln2}+. To cross-reference it, use the command \verb+\ref{algln2}+ for which it comes up as line~\ref{algln2} of Algorithm~\ref{algo1}.

% \subsection{Details on reference citations}\label{subsec7}

% Standard \LaTeX\ permits only numerical citations. To support both numerical and author-year citations this template uses \verb+natbib+ \LaTeX\ package. For style guidance please refer to the template user manual.

% Here is an example for \verb+\cite{...}+: \cite{bib1}. Another example for \verb+\citep{...}+: \citep{bib2}. For author-year citation mode, \verb+\cite{...}+ prints Jones et al. (1990) and \verb+\citep{...}+ prints (Jones et al., 1990).

% All cited bib entries are printed at the end of this article: \cite{bib3}, \cite{bib4}, \cite{bib5}, \cite{bib6}, \cite{bib7}, \cite{bib8}, \cite{bib9}, \cite{bib10}, \cite{bib11}, \cite{bib12} and \cite{bib13}.

% \section{Examples for theorem like environments}\label{sec10}

% For theorem like environments, we require \verb+amsthm+ package. There are three types of predefined theorem styles exists---\verb+thmstyleone+, \verb+thmstyletwo+ and \verb+thmstylethree+ 

% %%=============================================%%
% %% For presentation purpose, we have included  %%
% %% \bigskip command. Please ignore this.       %%
% %%=============================================%%
% \bigskip
% \begin{tabular}{|l|p{19pc}|}
% \hline
% \verb+thmstyleone+ & Numbered, theorem head in bold font and theorem text in italic style \\\hline
% \verb+thmstyletwo+ & Numbered, theorem head in roman font and theorem text in italic style \\\hline
% \verb+thmstylethree+ & Numbered, theorem head in bold font and theorem text in roman style \\\hline
% \end{tabular}
% \bigskip
% %%=============================================%%
% %% For presentation purpose, we have included  %%
% %% \bigskip command. Please ignore this.       %%
% %%=============================================%%

% For mathematics journals, theorem styles can be included as shown in the following examples:

% \begin{theorem}[Theorem subhead]\label{thm1}
% Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. 
% Example theorem text. Example theorem text. Example theorem text. Example theorem text. Example theorem text. 
% Example theorem text. 
% \end{theorem}

% Sample body text. Sample body text. Sample body text. Sample body text. Sample body text. Sample body text. Sample body text. Sample body text.

% \begin{proposition}
% Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. 
% Example proposition text. Example proposition text. Example proposition text. Example proposition text. Example proposition text. 
% \end{proposition}

% Sample body text. Sample body text. Sample body text. Sample body text. Sample body text. Sample body text. Sample body text. Sample body text.

% \begin{example}
% Phasellus adipiscing semper elit. Proin fermentum massa
% ac quam. Sed diam turpis, molestie vitae, placerat a, molestie nec, leo. Maecenas lacinia. Nam ipsum ligula, eleifend
% at, accumsan nec, suscipit a, ipsum. Morbi blandit ligula feugiat magna. Nunc eleifend consequat lorem. 
% \end{example}

% Sample body text. Sample body text. Sample body text. Sample body text. Sample body text. Sample body text. Sample body text. Sample body text.

% \begin{remark}
% Phasellus adipiscing semper elit. Proin fermentum massa
% ac quam. Sed diam turpis, molestie vitae, placerat a, molestie nec, leo. Maecenas lacinia. Nam ipsum ligula, eleifend
% at, accumsan nec, suscipit a, ipsum. Morbi blandit ligula feugiat magna. Nunc eleifend consequat lorem. 
% \end{remark}

% Sample body text. Sample body text. Sample body text. Sample body text. Sample body text. Sample body text. Sample body text. Sample body text.

% \begin{definition}[Definition sub head]
% Example definition text. Example definition text. Example definition text. Example definition text. Example definition text. Example definition text. Example definition text. Example definition text. 
% \end{definition}

% Additionally a predefined ``proof'' environment is available: \verb+\begin{proof}+ \verb+...+ \verb+\end{proof}+. This prints a ``Proof'' head in italic font style and the ``body text'' in roman font style with an open square at the end of each proof environment. 

% \begin{proof}
% Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. 
% \end{proof}

% Sample body text. Sample body text. Sample body text. Sample body text. Sample body text. Sample body text. Sample body text. Sample body text.

% \begin{proof}[Proof of Theorem~{\upshape\ref{thm1}}]
% Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. Example for proof text. 
% \end{proof}

% \noindent
% For a quote environment, use \verb+\begin{quote}...\end{quote}+
% \begin{quote}
% Quoted text example. Aliquam porttitor quam a lacus. Praesent vel arcu ut tortor cursus volutpat. In vitae pede quis diam bibendum placerat. Fusce elementum
% convallis neque. Sed dolor orci, scelerisque ac, dapibus nec, ultricies ut, mi. Duis nec dui quis leo sagittis commodo.
% \end{quote}

% Sample body text. Sample body text. Sample body text. Sample body text. Sample body text (refer Figure~\ref{fig1}). Sample body text. Sample body text. Sample body text (refer Table~\ref{tab3}). 

% \section{Methods}\label{sec11}

% Topical subheadings are allowed. Authors must ensure that their Methods section includes adequate experimental and characterization data necessary for others in the field to reproduce their work. Authors are encouraged to include RIIDs where appropriate. 

% \textbf{Ethical approval declarations} (only required where applicable) Any article reporting experiment/s carried out on (i)~live vertebrate (or higher invertebrates), (ii)~humans or (iii)~human samples must include an unambiguous statement within the methods section that meets the following requirements: 

% \begin{enumerate}[1.]
% \item Approval: a statement which confirms that all experimental protocols were approved by a named institutional and/or licensing committee. Please identify the approving body in the methods section

% \item Accordance: a statement explicitly saying that the methods were carried out in accordance with the relevant guidelines and regulations

% \item Informed consent (for experiments involving humans or human tissue samples): include a statement confirming that informed consent was obtained from all participants and/or their legal guardian/s
% \end{enumerate}

% If your manuscript includes potentially identifying patient/participant information, or if it describes human transplantation research, or if it reports results of a clinical trial then  additional information will be required. Please visit (\url{https://www.nature.com/nature-research/editorial-policies}) for Nature Portfolio journals, (\url{https://www.springer.com/gp/authors-editors/journal-author/journal-author-helpdesk/publishing-ethics/14214}) for Springer Nature journals, or (\url{https://www.biomedcentral.com/getpublished/editorial-policies\#ethics+and+consent}) for BMC.

% \section{Discussion}\label{sec12}

% Discussions should be brief and focused. In some disciplines use of Discussion or `Conclusion' is interchangeable. It is not mandatory to use both. Some journals prefer a section `Results and Discussion' followed by a section `Conclusion'. Please refer to Journal-level guidance for any specific requirements. 

% \section{Conclusion}\label{sec13}

% Conclusions may be used to restate your hypothesis or research question, restate your major findings, explain the relevance and the added value of your work, highlight any limitations of your study, describe future directions for research and recommendations. 

% In some disciplines use of Discussion or 'Conclusion' is interchangeable. It is not mandatory to use both. Please refer to Journal-level guidance for any specific requirements. 

% \backmatter

% \bmhead{Supplementary information}

% If your article has accompanying supplementary file/s please state so here. 

% Authors reporting data from electrophoretic gels and blots should supply the full unprocessed scans for key as part of their Supplementary information. This may be requested by the editorial team/s if it is missing.

% Please refer to Journal-level guidance for any specific requirements.

% \bmhead{Acknowledgements}

% Acknowledgements are not compulsory. Where included they should be brief. Grant or contribution numbers may be acknowledged.

% Please refer to Journal-level guidance for any specific requirements.

% \section*{Declarations}

% Some journals require declarations to be submitted in a standardised format. Please check the Instructions for Authors of the journal to which you are submitting to see if you need to complete this section. If yes, your manuscript must contain the following sections under the heading `Declarations':

% \begin{itemize}
% \item Funding
% \item Conflict of interest/Competing interests (check journal-specific guidelines for which heading to use)
% \item Ethics approval and consent to participate
% \item Consent for publication
% \item Data availability 
% \item Materials availability
% \item Code availability 
% \item Author contribution
% \end{itemize}

% \noindent
% If any of the sections are not relevant to your manuscript, please include the heading and write `Not applicable' for that section. 

% %%===================================================%%
% %% For presentation purpose, we have included        %%
% %% \bigskip command. Please ignore this.             %%
% %%===================================================%%
% \bigskip
% \begin{flushleft}%
% Editorial Policies for:

% \bigskip\noindent
% Springer journals and proceedings: \url{https://www.springer.com/gp/editorial-policies}

% \bigskip\noindent
% Nature Portfolio journals: \url{https://www.nature.com/nature-research/editorial-policies}

% \bigskip\noindent
% \textit{Scientific Reports}: \url{https://www.nature.com/srep/journal-policies/editorial-policies}

% \bigskip\noindent
% BMC journals: \url{https://www.biomedcentral.com/getpublished/editorial-policies}
% \end{flushleft}

% \begin{appendices}

% \section{Section title of first appendix}\label{secA1}

% An appendix contains supplementary information that is not an essential part of the text itself but which may be helpful in providing a more comprehensive understanding of the research problem or it is information that is too cumbersome to be included in the body of the paper.

% %%=============================================%%
% %% For submissions to Nature Portfolio Journals %%
% %% please use the heading ``Extended Data''.   %%
% %%=============================================%%

% %%=============================================================%%
% %% Sample for another appendix section			       %%
% %%=============================================================%%

% %% \section{Example of another appendix section}\label{secA2}%
% %% Appendices may be used for helpful, supporting or essential material that would otherwise 
% %% clutter, break up or be distracting to the text. Appendices can consist of sections, figures, 
% %% tables and equations etc.

% \end{appendices}

%%===========================================================================================%%
%% If you are submitting to one of the Nature Portfolio journals, using the eJP submission   %%
%% system, please include the references within the manuscript file itself. You may do this  %%
%% by copying the reference list from your .bbl file, paste it into the main manuscript .tex %%
%% file, and delete the associated \verb+\bibliography+ commands.                            %%
%%===========================================================================================%%

\bibliography{sn-bibliography}% common bib file
%% if required, the content of .bbl file can be included here once bbl is generated
%%\input sn-article.bbl

\end{document}