Abstract: Due to the inherent defects of infrared image, such as low contrast, blurred boundary and target background thermal confusion, the accuracy of traditional visual algorithm semantic understanding is insufficient. In this paper, the thermal radiation guided feature enhancement module (TAFEM) and multi-scale context fusion and edge refinement mechanism (MCFER) are proposed to solve the problem of small target loss and blurred boundary through the channel level thermal saliency weight dynamic modulation feature, combined with cross layer attention pyramid and edge monitoring branch; The IR-Yolo architecture is proposed to realize the detection segmentation collaborative semantic reading. Experimental results demonstrate that the proposed method improves the mAP@0.5 to 0.842 in nighttime scenarios, which is the best performance in the mainstream model, and verifies the effectiveness of the method.