Global-Regularized Neighborhood Regression for Efficient Zero-Shot Texture Anomaly Detection

TL;DR

This work addresses the challenge of texture anomaly detection under open-set, zero-shot conditions by proposing Global-Regularized Neighborhood Regression (GRNR). GRNR extracts two intrinsic priors—the local neighborhood prior $S_L$ and the global normality prior $S_G$—directly from the test image and uses them in a self-reconstructive regression objective with a fast closed-form solution. The approach achieves competitive or superior performance across defect-close, texture-close, and open-set tasks without any training data, underscoring its data- and labor-efficiency. By demonstrating strong results on the Texture Spectrum benchmark and providing detailed ablations, GRNR offers a practical, training-free pathway for industrial texture anomaly detection with open-set capabilities.

Abstract

Texture surface anomaly detection finds widespread applications in industrial settings. However, existing methods often necessitate gathering numerous samples for model training. Moreover, they predominantly operate within a close-set detection framework, limiting their ability to identify anomalies beyond the training dataset. To tackle these challenges, this paper introduces a novel zero-shot texture anomaly detection method named Global-Regularized Neighborhood Regression (GRNR). Unlike conventional approaches, GRNR can detect anomalies on arbitrary textured surfaces without any training data or cost. Drawing from human visual cognition, GRNR derives two intrinsic prior supports directly from the test texture image: local neighborhood priors characterized by coherent similarities and global normality priors featuring typical normal patterns. The fundamental principle of GRNR involves utilizing the two extracted intrinsic support priors for self-reconstructive regression of the query sample. This process employs the transformation facilitated by local neighbor support while being regularized by global normality support, aiming to not only achieve visually consistent reconstruction results but also preserve normality properties. We validate the effectiveness of GRNR across various industrial scenarios using eight benchmark datasets, demonstrating its superior detection performance without the need for training data. Remarkably, our method is applicable for open-set texture defect detection and can even surpass existing vanilla approaches that require extensive training.

Figures (9)

• Figure 1: A comprehensive comparison of existing textured surface anomaly defect detection methods. (a) Supervised learning methods necessitate a substantial amount of labeled data for training and operate within a defect close-set detection paradigm. This means they can only identify defect types present in the training dataset. (b)-(c). Unsupervised learning and few-shot learning methods require normal samples during model training and operate within a texture close-set framework. They are limited to detecting texture types encountered in the training dataset. (d) The proposed zero-shot learning method stands out by not requiring any training cost. It operates under an open-set detection paradigm, enabling it to identify texture defects in open-world industrial scenarios.
• Figure 2: Method pipeline. The image is converted into patch-level embeddings using a feature extractor. Subsequently, global and local neighbor support elements $\left \{ \bm{S}_L, \bm{S}_G \right \}$, are derived through self-filtering and neighborhood operations. Following this, the global-regularized neighborhood regression process is executed to reconstruct individual query elements $\bm{Q}$. The optimal transformation $\tilde{\bm{W}}$ considers that the obtained $\bar{\bm{Q}}$ through $\tilde{\bm{W}}\bm{S}_L$ can achieve similarity with $\bm{Q}$(regression direction) while simultaneously adhering to the global normal distribution defined by $\bm{S}_G$(regularization direction). Anomaly scores are obtained through a feature-level comparison.
• Figure 3: Some representative samples from the constructed Texture Spectrum dataset, which encompasses a diverse array of textured surfaces featuring both homogeneous and non-homogeneous textures. Moreover, it includes various defect types such as cracks, holes, scratches, and others.
• Figure 4: (a). Printing product production site. (b). Our AOI instruments for printing defect inspection.
• Figure 5: The detailed configuration of each sub-dataset in Texture Spectrum, $N_t$ represents the number of texture types.