DFA-YOLO:Accuracy Enhancement Method for Non-Rigid Object Detection under Complex Conditions

In industrial safety and disaster monitoring, non-rigid objects such as smoke and flame pose significant challenges for detection due to their rapidly changing shapes and scales over time as well as their blurred boundaries. We present DFA-YOLO, a customized YOLOv11-based detector tailored for deformable targets. The framework couples DCNv4 to adaptively model local deformations, integrates ACmix to fuse global dependencies with local details, and adopts Wise-IoU (WIoU) to stabilize bounding-box regression for irregular shapes. Evaluations on three representative datasets—VOC (general scenes), Smoke-Fire (highly deformable targets), and Elec (industrial arc discharges)—show higher accuracy at competitive throughput over mainstream YOLO baselines while maintaining competitive throughput. In particular, DFA-YOLO improves precision/robustness on de-formable targets and achieves state-of-the-art accuracy–speed trade-offs on the industrial dataset. These results indicate that DFA-YOLO provides a practical solution for safety monitoring and early-warning applications in real-world environments.