Abstract: Owing to the inherent defects of infrared images, such as their low contrast, blurred boundaries, and target background thermal confusion, traditional visual algorithms provide an insufficient semantic understanding. This study investigated a thermal radiation-guided feature-enhancement module and multi-scale context-fusion and edge-refinement mechanism to solve the small-target loss and blurred-boundary problems. A channel-level thermal saliency weight dynamic modulation feature was combined with a cross-layer attention pyramid and edge-monitoring branch. The IR-Yolo architecture was used to realize detection segmentation collaborative semantic reading. Experimental results demonstrated that the proposed method improved the mAP@0.5 metric, with a value of 0.842 in nighttime scenarios, which was the best performance obtained for a mainstream model, and verified the effectiveness of the method.