Optimized SAR Image Recognition Algorithm Based on Lightweight YOLO Network

To address the problems of parameter redundancy, high computational complexity, and difficulty in embedded deployment of Synthetic Aperture Radar (SAR) image target detection models, this paper proposes a lightweight recognition algorithm capable of suppressing environmental interference. The algorithm reduces computational complexity through depthwise separable convolution, achieves parameter compression via a framework integrating quantization and sparse pruning, enhances robustness in complex scenes with an adaptive prior knowledgeassisted soft activation mask, and designs a module-level error decoupling and traceability mechanism to realize fault localization and interpretability improvement of the compressed model. Experimental results show that the model trained on vertically polarized 2D features achieves an average accuracy of 86.3 % across four types of scenes and an overall accuracy of 8 4 %. It maintains strong adaptability to rotated, occluded, and atypical targets, and can accurately locate the root cause of model performance degradation. While reducing complexity, the algorithm balances detection accuracy, anti-interference capability, and interpretability.