Rethinking interactive networks and regression loss functions for industrial defect detection

Abstract

Defect detection of industrial products is crucial to ensure product quality and use safety. However, the following difficulties still exist complex defect background, the large similarity between defect classes and weak semantic information, and extensive image noise information. In order to solve this challenge above, this paper proposes a novel Interactive Convolutional Transformer-based Encoder-Decoder Defect Detection Network (ICT-EDNet). Specifically, ICT-EDNet has the following three characteristics. First, this paper designs the edge-interactive deep convolution (EIDC) module and the feature cyclic shift transformer (FCST) module. It can both reduce the interference of noise information and effectively capture the effective semantic relationship between arbitrary positions, which is committed to strengthening the semantic information of key features and defects. Secondly, considering the feature differences between EIDC and FCST, this paper designs the Feature Similarity Bridging (FSB) module and Feedback Spatial Information Regulation (FSIR) feedback mechanism. It can better ensure the local details of the global representation and the global perceptibility of the local features. Finally, this paper proposes the perimeter and aspect ratio complementary adaptive tuning segmentation function ZCIoU, which accelerates the convergence of the model. The experimental results show that ICT-EDNet outperforms the recently proposed SOTA in NEU, DeepPCB, and Track Components datasets.