A dual-branch attention-based segmentation model for accurate detection of small-scale terminal surface defects in industrial environments

Abstract

Terminals are components used to link cables in electronic systems. Burn defects that may occur on the surfaces of terminals during production can cause performance losses in energy transmission systems. This study proposes a novel method to overcome the limitations of manual inspection in post-production quality control. The proposed segmentation model, named DFA-AttenNet, is enhanced with Dilated Feature Aggregation (DFA) and attention mechanisms, and is specifically designed to detect low-contrast and small-scale defects with high accuracy. A large dataset was created and labeled using an image acquisition system, and the DFA-AttenNet model was tested on this dataset. The experiments conducted in this study demonstrate that the proposed model is suitable for real production lines, due to its capability for detailed defect analysis and high accuracy across different datasets. In addition, with the embedded system we installed in the production line, when a defective product arrives and is detected, it stops the system and warns the operator. This ensures that the defective item is removed from the line without interrupting production, and prevents it from being shipped to the customer. This study aims to provide a reliable and scalable defect detection model by increasing automation in industrial quality control processes.