Continuous casting billet defect detection based on improved YOLOv5

During the production of continuous cast steel, under the constraints of various factors such as material qualities and processing techniques, various types of defects may occur within the steel, such as inter-cracks and shrinkage holes, which will affect the quality of the steel. Continuous casting billet defect detection is different from the general target detection task, which involves small targets and targets with very high aspect ratios. To address the target characteristics, this paper uses improvements based on the YOLOv5 algorithm to improve the accuracy of continuous casting billet defect detection. Firstly, the self-constructed dataset is analysed, while a clustering algorithm based on differential evolutionary algorithm is applied to re-cluster the anchor frames of small targets and extreme aspect ratio targets. For small targets, the attention mechanism and BiFPN structure are used to improve feature extraction. Finally, varifocal loss is introduced to mitigate the effect of positive and negative sample ratio imbalance on the training results. The experimental results show that the mAP of the improved algorithm is improved by 5.1% compared to the original YOLOv5.