Abstract: To address the problems of blurred boundaries of hot-spot targets,complex temperature gradient variations,and large differences in target scales in infrared image object detection for electrical equipment in mining-area substations,existing detection algorithms generally have deficiencies such as insufficient recognition accuracy and high missed detection rates.By comparing the performance of several mainstream object detection algorithms on an infrared image dataset of electrical equip ment in mining-area substations,it is found that the Faster R-CNN algorithm incorporating a feature pyramid structure performs relatively well,but its recognition performance of small hot-spot targets and temperature anomaly regions still has room for im provement.Therefore,an improved Faster R-CNN algorithm is proposed.First,a spatial pyramid pooling structure is introduced into the network to extract multi-scale features.Second,the feature fusion method is optimized to enhance the network′s percep tion of targets at different scales.In the region proposal generation stage,the quality of region proposals is improved by expan ding anchor scales and aspect ratios.Finally,the classification and bounding box regression loss functions are improved to en hance detection accuracy.Experimental results show that,compared with the original Faster R-CNN algorithm,the improved al gorithm achieves improvements of 2.7 percentage points in AP(0.5∶0.95),4.3 percentage points in AP(0.5),and 1.8 per centage points in the detection accuracy for small-sized hot-spot targets on the infrared image dataset of electrical equipment in mining-area substations,reflecting that the improved algorithm has certain applicability to infrared detection of mining-area e lectrical equipment.

Key words: mining area substation,infrared image,electrical equipment,object detection,Faster R-CNN,multi-scale fea ture fusion