Abstract: The paper proposes a lightweight structured pruning framework tailored for anti-drone scenarios. The framework first employs structural decoupling transformations to equivalently convert highly coupled complex feature extraction modules in YOLO26(i.e., C2, C2f and C3k2) into channel-independent regularization operators, laying the architectural foundation for subsequent pruning; it then integrates three importance evaluation strategies—Random, L1, and LAMP—to assign quantitative scores to convolutional channels, and performs progressive iterative pruning under the constraint of a target acceleration ratio, gradually removing redundant channels through multiple steps of small-scale compression; finally, its accuracy is restored through fine-tuning. Experimental results show that the pruned model requires only 1.2 million parameters, a reduction of nearly 50% compared to the mainstream lightweight model YOLO26, and achieves an inference speed of 1 355.7 FPS, the highest among all compared models. In terms of accuracy, the model's mAP50 is 0.965, differing from YOLO26's 0.966 by only 0.001; its mAP50-95 is 0.652, also second only to YOLO26. Compared to models such as YOLOv5, YOLOv8, YOLOv11, and YOLOv12, this model demonstrates significant advantages in both accuracy and speed despite a substantial reduction in the number of parameters, which verifies the effectiveness and practicality of the proposed pruning method.