A Parallel Feature Extraction Model with Channel Attention for Button Defect Detection

Defect detection is a vital task for production process, which greatly affects the quality of the product. Manual inspection is the most commonly used defect detection method, but is very time-consuming and not reliable. Currently, many image processing algorithms have been used for defect detection to partially replace manual detection. However, due to the complexity of various defects, how to correctly and quickly detect whether the button is defective still faces great challenges. To address these challenges, we propose a button defect detection model based on parallel feature extraction network. The model extracts feature by two parallel CNN structures, one of which is deeper and stronger (CNN-B), and the other is relatively shallow and weak (CNN-A). The CNN-B is trained on ImageNet firstly. Meanwhile, only the three layers of CNN-B but all layers in CNN-A are trained. We further add channel attention module to perform feature maps recalibration in CNN-A. The experimental results show that the parallel feature extraction model can identify the defect of button effectively.