Inception meets Swin Transformer: A Novel Approach for Metal Defect Recognition

Abstract

The detection of metal defects with high precision and efficiency is a significant challenge in modern industry. Existing machine learning methods for recognizing common metal surface defects heavily rely on expert knowledge for manual feature extraction. Conventional deep learning methods face challenges in capturing global feature information from defect images or defect detection signals.To address this issue, we proposed a metal defect recognition method based on an Inception-fused Swin Transformer model. The method combines the adaptive local feature extraction capability of the Inception structure with the advantage of the Swin Transformer in capturing global feature information from defect signals. Additionally, it utilizes the Channel-Coordinate Attention module (CoordAttention) to highlight important feature channels. Experimental results demonstrate the effectiveness of the proposed method on the Ultrasonic Defect Grayscale Image dataset (ULFSL-DET) and the publicly available Image-based Metal Defect dataset (NEU-CLS), achieving recognition accuracies of 98.1% and 99.8%, respectively. The method exhibits high effectiveness in recognizing metal defect signals in grayscale images, and it demonstrates strong generality for image-based metal defect recognition.