Randomized-enhanced Magnitude Channel Pruning for Radar Jamming Detection

With the rapid development of radar jamming technology, traditional detection techniques face significant limitations in complex electromagnetic environments. Recent interference detection methods based on deep learning struggle with processing speed and real-time performance. To address this, we propose a new pruning technique called randomized-enhanced magnitude channel pruning (RMP) to achieve faster detection speeds and more lightweight models. This study evaluates the performance of RMP compared to random channel pruning and magnitude channel pruning in radar jamming detection. To verify RMP's generalization, we also tested it on the VOC2007 [16] object detection task. Results show that RMP consistently reduces mAP@50 loss more effectively than other methods. In one iteration pruning, RMP reduced mAP@50 loss by an average of 0.8% and 1% in radar jamming detection, and by 0.4% and 3.1% in the VOC2007 dataset, compared to random and magnitude pruning, respectively. In multi-pass pruning, RMP reduced mAP@50 loss by an average of 0.3% in radar jamming detection, and by 0.7% and 1.6% in the VOC2007 dataset, compared to random and magnitude pruning, respectively. These findings confirm RMP's effectiveness in maintaining model accuracy and its practical value in fast-response scenarios like radar jamming detection.