Steel Surface Defect Detection Method Based on Improved YOLOv4-tiny

Abstract

As small-scale targets on steel surfaces are prone to miss detection and the current mainstream target detection algorithms have the problems of large number of parameters and computational complexity, an improved YOLOv4-tiny method for steel surface defect detection is proposed. The CARAFE lightweight universal upsampling operator is used instead of upsampling to improve the feature fusion capability; the residual feature enhancement structure ASPP_ASF+ is introduced to capture multi-scale contextual information by parallel sampling with the null convolution of different expansion coefficients, and then the feature weighted fusion is realized by the ASF+ component to improve the detection accuracy of small targets; the Mish activation function is used to increase the nonlinear expression and enhance the network generalization capability. In the NEU_DET dataset, the improved YOLOv4-tiny network model improves the mAP by 6.17% compared with the original model; in the VOC2007 dataset, the mAP improves by 3.37%. The experimental results show that the improved algorithm can effectively improve the detection ability of the model and is generalizable.