Rethinking Mask Down Sampling and Regression Loss Function in Industrial Tiny Defect Detection

PCB is an important precision component, and its production process will have tiny short circuits, mouse bites, and other defects. In tiny object detection, bounding box regression and image down sampling are the key factors to determine the object detection performance. However, the previous loss function and down sampling methods have disadvantages such as inaccurate localization, which leads to slow convergence of defect detection and inaccurate detection results. In this paper, we improve the loss function and down sampling methods to perform more accurate industrial tiny defect detection. Firstly, we propose the TDIOU loss function based on the area loss. Secondly, a dynamic mask-based down sampling method is proposed to help the detector automatically filter important features, minor features, and noisy features in the process of reducing the size of the parameter matrix. Finally, a one-stage defect detector based on the above-mask down sampling and TDIOU loss function is built, and comparison experiments and ablation experiments are conducted on the Peking University PCB defect dataset and Deep PCB dataset to verify the advantages of this method.