Semi-supervised Knowledge Distillation for Tiny Defect Detection

Image anomaly detection can automatically detect defects using images of products, which is crucial for product quality controls. Because of insufficient abnormal data, unsupervised image anomaly detection based on knowledge distillation has attracted broad attention recently. However, fully unsuper...

Full description

Saved in:
Bibliographic Details
Published in2022 IEEE 25th International Conference on Computer Supported Cooperative Work in Design (CSCWD) pp. 1010 - 1015
Main Authors Cao, Yunkang, Song, Yanan, Xu, Xiaohao, Li, Shuya, Yu, Yuhao, Zhang, Yifeng, Shen, Weiming
Format Conference Proceeding
LanguageEnglish
Published IEEE 04.05.2022
Subjects
Collaborative work
Conferences
Convolutional Neural Network
Feature extraction
Filling
Image Anomaly Detection
Knowledge Distillation
Neural networks
Pre-trained Network
Product design
Quality assessment
Online AccessGet full text

Cover

More Information
Summary:Image anomaly detection can automatically detect defects using images of products, which is crucial for product quality controls. Because of insufficient abnormal data, unsupervised image anomaly detection based on knowledge distillation has attracted broad attention recently. However, fully unsupervised methods suffer from detecting tiny anomalies that widely exist in industrial products because the features of tiny anomalies and normal features extracted by the teacher network are similar. This paper extends current unsupervised anomaly detection methods into a semi-supervised manner, simultaneously leveraging normal data and a limited amount of abnormal data. An automobile plastic parts dataset is established to prove the effectiveness of the proposed method. Experiments show that the proposed method can accurately detect small anomalies and largely surpass a powerful baseline (6% in AU-ROC, 10% in F1-score, 11% in Accuracy).
DOI:10.1109/CSCWD54268.2022.9776026