Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Distillation-based fabric anomaly detection

Simon Thomine, Hichem Snoussi

TL;DR

The paper tackles unsupervised fabric anomaly detection in highly variable textures by introducing a residual reverse-distillation framework that emphasizes high-level texture features while restricting the student from reconstructing anomalies. It incorporates a bottleneck embedding (1×1 convolutions with a SSPCAB attention block) and residual connections to selectively transfer informative features from a frozen teacher to a lighter student, plus a domain-generalization variant that omits residuals to better capture deeper descriptors. Key contributions include a robust texture anomaly detector using reverse knowledge distillation, a domain-generalized variant for fabrication environments, and a newly collected industrial textile dataset to evaluate real-world performance. Experiments show strong AUROC on fabric-like textures (MVTEC AD, AITEX, TILDA) and impressive inference speeds, supporting practical deployment in online textile inspection and quality control. The work advances industrial anomaly detection by combining efficient architectures with targeted feature selection and provides a valuable dataset for ongoing research.

Abstract

Unsupervised texture anomaly detection has been a concerning topic in a vast amount of industrial processes. Patterned textures inspection, particularly in the context of fabric defect detection, is indeed a widely encountered use case. This task involves handling a diverse spectrum of colors and textile types, encompassing a wide range of fabrics. Given the extensive variability in colors, textures, and defect types, fabric defect detection poses a complex and challenging problem in the field of patterned textures inspection. In this article, we propose a knowledge distillation-based approach tailored specifically for addressing the challenge of unsupervised anomaly detection in textures resembling fabrics. Our method aims to redefine the recently introduced reverse distillation approach, which advocates for an encoder-decoder design to mitigate classifier bias and to prevent the student from reconstructing anomalies. In this study, we present a new reverse distillation technique for the specific task of fabric defect detection. Our approach involves a meticulous design selection that strategically highlights high-level features. To demonstrate the capabilities of our approach both in terms of performance and inference speed, we conducted a series of experiments on multiple texture datasets, including MVTEC AD, AITEX, and TILDA, alongside conducting experiments on a dataset acquired from a textile manufacturing facility. The main contributions of this paper are the following: a robust texture anomaly detector utilizing a reverse knowledge-distillation technique suitable for both anomaly detection and domain generalization and a novel dataset encompassing a diverse range of fabrics and defects.

Distillation-based fabric anomaly detection

TL;DR

The paper tackles unsupervised fabric anomaly detection in highly variable textures by introducing a residual reverse-distillation framework that emphasizes high-level texture features while restricting the student from reconstructing anomalies. It incorporates a bottleneck embedding (1×1 convolutions with a SSPCAB attention block) and residual connections to selectively transfer informative features from a frozen teacher to a lighter student, plus a domain-generalization variant that omits residuals to better capture deeper descriptors. Key contributions include a robust texture anomaly detector using reverse knowledge distillation, a domain-generalized variant for fabrication environments, and a newly collected industrial textile dataset to evaluate real-world performance. Experiments show strong AUROC on fabric-like textures (MVTEC AD, AITEX, TILDA) and impressive inference speeds, supporting practical deployment in online textile inspection and quality control. The work advances industrial anomaly detection by combining efficient architectures with targeted feature selection and provides a valuable dataset for ongoing research.

Abstract

Unsupervised texture anomaly detection has been a concerning topic in a vast amount of industrial processes. Patterned textures inspection, particularly in the context of fabric defect detection, is indeed a widely encountered use case. This task involves handling a diverse spectrum of colors and textile types, encompassing a wide range of fabrics. Given the extensive variability in colors, textures, and defect types, fabric defect detection poses a complex and challenging problem in the field of patterned textures inspection. In this article, we propose a knowledge distillation-based approach tailored specifically for addressing the challenge of unsupervised anomaly detection in textures resembling fabrics. Our method aims to redefine the recently introduced reverse distillation approach, which advocates for an encoder-decoder design to mitigate classifier bias and to prevent the student from reconstructing anomalies. In this study, we present a new reverse distillation technique for the specific task of fabric defect detection. Our approach involves a meticulous design selection that strategically highlights high-level features. To demonstrate the capabilities of our approach both in terms of performance and inference speed, we conducted a series of experiments on multiple texture datasets, including MVTEC AD, AITEX, and TILDA, alongside conducting experiments on a dataset acquired from a textile manufacturing facility. The main contributions of this paper are the following: a robust texture anomaly detector utilizing a reverse knowledge-distillation technique suitable for both anomaly detection and domain generalization and a novel dataset encompassing a diverse range of fabrics and defects.
Paper Structure (17 sections, 4 equations, 9 figures, 9 tables)

This paper contains 17 sections, 4 equations, 9 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Residual Reverse distillation using specifically selected embeddings
  4. Residual reverse distillation
  5. Bottleneck Embedding
  6. Loss and Anomaly Score Calculation
  7. Reverse distillation for domain generalization
  8. Industrial textile dataset
  9. Industry constraints
  10. Variety of defects
  11. Dataset description
  12. Existing datasets
  13. Experiments
  14. State of the art AUROC comparison
  15. Inference speed comparison
  16. ...and 2 more sections

Figures (9)

  • Figure 1: Attention module used for residual connection
  • Figure 2: Reverse residual distillation architecture.
  • Figure 3: Feature fusion and bottleneck architecture
  • Figure 4: Images extracted from our proposed dataset with defect that could be dust
  • Figure 5: Images of our proposed datasets, (upper) defect-free images, (lower) defective images. Some images are a bit blurry due to motion blur. Other kind of defective samples are presented in the experiment part.
  • ...and 4 more figures