The deep detection model based on the dual-branch architecture has shown significant efficiency advantages in real-time tasks, its inherent feature coupling problem often leads to the semantic dilution of fine-grained features. Especially in high-complexity industrial scenarios, the detailed information is susceptible to the interference of contextual features, which limits the further improvement of detection accuracy. To address this challenge, this paper aims to propose a novel and real-time three-branch network fusing edge prior knowledge network (EPKTBNet) for visual defect detection.
The EPKTBNet is proposed, which contains three branches, namely, detail branch, context branch and edge branch, which are used to parse detail, context and edge information respectively. The edge branch combines the existing edge knowledge to identify high-frequency edge features and predict the boundary area, which effectively improves the positioning accuracy of the defect boundary.
Some experimental results show that EPKTBNet is superior to the existing state-of-the-art methods on multiple industrial benchmark data sets, especially in the defect boundary segmentation task, achieving the best trade-off between speed and accuracy. Specifically, EPKTBNet achieves 51.6% mIOU with inference speed of 145.5 FPS on ScrewDefect data set and 66.8% mIOU with speed of 197.2 FPS on CrackForest data set.
By introducing the three-branch network fusing edge prior knowledge, this paper solves the problem that fine-grained features are susceptible to interference in industrial defect detection, provides new ideas for high-precision real-time detection in complex scenes and promotes the development of industrial visual inspection theory and application.
