A light deformable multi-scale defect detection model for irregular small defects with complex background

Abstract

Surface defect detection is a crucial aspect of quality inspection in the industrial sector, hindered by the challenges of recognizing irregular small defects and dealing with complex background interference. To address these issues, we proposed a novel model called Deformable Efficient Multi-Scale Net for Small Defects (DSE-NET), which incorporates three innovative components: (1) The Inverted Residual Efficient Multi-Scale Attention (iREMA) alleviates complex background interference issues efficiently through regions of interest. (2) The Small Defect Feature Pyramid Network (SFPN) addresses the issue of small defects by progressively processing the added small target feature layer. (3) The Deformable Darknet mitigates irregular defect problems through deformable convolutions. Extensive ablation and comparative experiments were conducted on our self-built Micro Bearing Defect database(MB-DET), Northeastern University Detection (NEU-DET) database, and Peking University Printed Circuit Board Detection (PCB-DET) dataset. Compared to the baseline model, DSE-NET has increased the Mean Average Precision at IoU = 0.5 ([email protected]) accuracy by 2.9 % ∼ 5.2 % while only expanding the Floating Point Operations per Second (FLOPs) by 1.2G and the number of parameters by 0.3 M. The work has contributed various efficient components in the Artificial Intelligence field and effectively alleviated various thorny issues in industrial surface defect detection. In addition, we developed a surface defect detection system based on real industrial scenarios and implemented the mobile deployment of the detection model, which verified the competitiveness of DSE-Net real-time detection. The relevant model deployment method code is available: https://github.com/17854222655/Mobile-Deployment.git.