Abstract: To solve the problems of low accuracy and the poor effect of mainstream object detection algorithms on infrared-visible fusion images, a target-detection method based on improved YOLOv8-s is proposed. First, PSA and ACmix were used to create convolutional, multi-scale fusion, and self-attention parallel structures in YOLOv8-s, which were integrated into the Conv convolutional layer to better capture long-distance dependencies under lower model parameters. Second, a multi-channel upsampling cascade fusion module dominated by DySample was added to increase the sampling richness and enhance the information extraction between channels to improve the target recognition accuracy and the coincidence of the positioning frame. Finally, the auxiliary learning module PGIv8, which was designed with reference to the PGI structure and used only during training, obtained better model parameters and detection results without increasing the inference cost. The results show that mAP@50 reached 98.4% and 81.1%, and mAP@50:95 reached 71.1% and 44.3%, respectively, and their average accuracies were 1.3% and 4.3% higher than that of the original YOLOv8-s, respectively, which meets the experimental expectations.