Improvements to YOLOv4 for Steel Surface Defect Detection

Abstract

The effective detection of surface defects in steel is a key concern for steel producers. Conventional target detection algorithms are unable to effectively detect defective targets. Using the advantages of deep learning in feature learning, a surface defect detection algorithm with improved YOLOv4 is proposed to address the problem of poor detection accuracy caused by a variety of steel surface defects and the presence of a large number of small areas and blurred edge damage. Firstly, TBConv (Tied Block Convolution) was introduced to improve the standard convolutional layer in the backbone feature extraction network CSPDarknet to enhance its feature learning capability for different types of surface defects. Secondly, the ECA(Efficient Channel Attention) attention mechanism is introduced in the detection layer to increase the weight of useful features while suppressing the weight of invalid features to improve the ability to fuse shallow and deep information. A cascading bottleneck residual structure is added after the SPP feature pyramid module, with the output of the previous stage being used as the input to the next stage in a sequential training process to enhance the target feature representation. The experimental results show that the improved YOLOv4 algorithm has 6.2% higher accuracy and 27.46% smaller model size compared to the original algorithm, and the algorithm has improved the detection rate of small area defects and can effectively detect surface defect targets.