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

Aiming at the PCB in the production process will appear short circuit, missing holes and other defects, the existing loss function and down sampling algorithms have inaccurate localization and so on leading to the slow convergence of defect detection, inaccurate detection results, a mask down sampling and tiny defects based on the intersection of the ratio of the loss of the one-stage defect detector is proposed. First, an area loss-based loss function is proposed for more accurate regression localization of tiny defects. Second, a dynamic mask-based down sampling algorithm is proposed to facilitate the automatic screening of important, secondary and noisy features in the process of reducing the size of the parameter matrix, so as to improve the feature extraction capability of the defect detector. The experimental results show that the proposed defect detector can achieve 98.50% mAP and 99.02% mAP on the Peking University PCB defect dataset and Deep PCB dataset, respectively, which is better than the comparison algorithms; the proposed downsampling algorithm improves the mAP of the detectors such as YOLOv5 in the PCB dataset of Peking University by 1.6 percentage points. The proposed loss function facilitates the detectors such as YOLOv3 and YOLOv4 to improve their detection accuracies on the two datasets, showing good robustness.