A lightweight and efficient model for surface tiny defect detection.

Due to the tiny size and the fuzzy pixels of tiny objects, tiny defect detection is a thorny problem in industry application. Meanwhile, the real-time detection of tiny defects is highly required in industrial assembly line. That is to say, tiny defect detection algorithms must ensure high accuracy while maintaining a low computational cost. This paper designs a lightweight and efficient network for tiny defect detection. In the backbone network, Diagonal Feature Pyramid (DFP) is proposed to improve the performance of tiny defect detection. For higher accuracy, DFP fuses more original features if they are at the same level. For less computational cost, DFP reduces the model size by eliminating the bottom-up pathway and removing some non-original same-level features. In the neck network, a multi-scale neck network with some fusion strategies is designed to suit multi-scale tiny defect detection. Finally, an adaptive localization loss function is designed to adjust the sensitivity of tiny defects. Based on a public PCB (Printed Circuit Board) dataset, the comparative experiments show that our model has better mAP and higher speed than various mainstream defect detection algorithms. [ABSTRACT FROM AUTHOR]