Vehicle-mounted infrared target detection based on improved YOLOv5 model

To address the issue of low detection accuracy in vehicle-mounted infrared target detection algorithms, caused by the limited features and low contrast of small targets, this paper proposed the FE-YOLOv5s vehicle-mounted infrared target detection model. Firstly, the PPA module was integrated into the backbone network to mitigate the feature information loss caused by downsampling operations. Secondly, based on the AFPN network and the DASI module, we developed a feature fusion network termed DAFPN. This network directly fuses feature information from non-adjacent scale layers and incorporates features from five different scale layers into the final feature output, further enhancing the feature representation capability of the targets. The target prediction layer was reconstructed to improve the model's recognition rate of small targets. Finally, MPDIoU was adopted as the bounding box loss function to enhance target localization accuracy. Experimental results demonstrated that the proposed model achieves mAP@0.5 scores of 88.1% on the FLIR dataset and 84.7% on the iRay vehicle-mounted infrared dataset, representing improvements of 7.2% and 3.2% over the YOLOv5s model. It effectively alleviates the problem of false and missed detections of infrared small targets while operating at 100 FPS, which confirms its potential for embedded deployment.