Target temperature region detection of converter thermal infrared image based on improved YOLOv5s

Abstract

Aiming at the difficulty of real-time temperature detection in the converter smelting process, most of the production sites use sub-guns for only end point detection, In this paper, the YOLOv5s-XCB detection algorithm is used to automatically extract the target temperature area of the converter thermal infrared image. It lays the foundation for the next step to realize automatic temperature measurement combined with the temperature matrix of this area. Based on the YOLOv5s algorithm, the research adds a small target detection layer and a CBAM attention mechanism to solve the problem that small targets and weak target temperature regions are difficult to detect. The BiFPN structure is used in the Neck layer to fuse the original feature information extracted by the backbone network to enhance the detection accuracy. The results show that the average mean precision (mAP) of the improved algorithm is 95.8%, the FPS is 69.5, and the confidence of the detection frame is significantly improved, which solves the problem that the original YOLOv5s algorithm is difficult to detect small target temperature areas and weak target temperature areas.