MFAF-YOLO: Vehicle Over-temperature Region Detection based on Multi-scale Feature Attention Fusion

Abstract Due to the similarity and scale diversity between foreground and background information of infrared images, missed and false detection easily occurs in high-speed vehicle over-temperature region detection. Aimed at this problem, based on the YOLOv10 framework, a multi-scale feature attention fusion improved YOLO (MFAF-YOLO) is proposed. Firstly, a spatially enhanced efficient multi-scale attention module is designed in the bottleneck layer of YOLOv10, combined with the pixel-level attention mechanism, the sensitive feature extraction capacity of the proposed MFAF-YOLO is effectively improved. Then, a feature filtering and cross-scale connected feature pyramid network (FFC-FPN) is proposed, which guides the filtering and screening of low-level feature information through high-level features to improve the discrimination of features. Meanwhile, by cross-scale feature weighted fusion, the multi-scale feature expression capacity of the proposed MFAF-YOLO is further enhanced. Finally, the results show that the mean Average Precision (mAP) indexes of the proposed method on the two datasets reach 97.2% and 97.8%, which is at least 3.9% higher than the existing detection methods, and has good comprehensive performance in the tire temperature detection of high-speed vehicles.