Auto Parts Defect Detection Based on Few-shot Learning

Abstract

In the process of auto parts products from manufacturing to quality inspection, it is inevitable that there will be defects on the surface due to aging of machine tools or human factors. How to better apply computer vision technology to defect detection of auto parts is still an issue discussed by the industry and academia. For the condition that sufficient labeled samples is usually not available in the industrial field, this paper uses an improved few-shot metric learning model to identify auto parts samples. According to the traditional few-shot learning models lack of the considerations of channel level and internal relations of support set, this paper adds ECA module and Transformer module to few-shot learning. Add more attention to the key information of support set and query set samples at channel level while increasing the attention to the intrinsic feature of support set, to improve the classification effect.