NOUFERROUS METALLURGICAL EQUIPMENT

This article proposes a high-resolution feature detection method for steel surface defects based on HRNet (High Resolution Net) to address the limited practical application and research of semantic segmentation algorithms in steel defect detection, as well as the low accuracy of traditional semantic segmentation methods in defect segmentation. The method adopts a multi-level network parallel framework to maintain high-resolution features, extracts high-resolution features with strong position sensitivity through parallel multi-resolution convolution and repeated multi-resolution fusion mechanisms, and combines HRNet to perform freeze training and migration learning on the enhanced dataset, ultimately achieving high-precision segmentation of steel surface defect images. The experiment shows that this method has a good segmentation effect on steel defects, with an average intersection to union ratio (MIOU) index and an average pixel accuracy (MPA) index of 94.92% and 97.64%, respectively.