MTKD-Det: An Efficient SAR Target Detector Based on Multi-Teacher Knowledge Distillation

As a critical branch of synthetic aperture radar (SAR) interpretation, SAR target detection has many applications in the military and civil fields. Although traditional deep learning methods achieve some satisfactory performance in target detection tasks, they inherently require significant computational resources and interpretation time. Hence, traditional deep learning methods frequently struggle to maintain a balance between high efficiency and precision. It is undoubtedly a huge challenge for target detection tasks. To address this problem, an efficient SAR target detector based on multi-teacher knowledge distillation (MTKD-Det) is proposed in this paper, which can effectively enhance target detection accuracy while reducing inference time. Specifically, a multiple-teacher knowledge distillation strategy is used in the proposed method, which integrates the experiences and methodologies of multiple teachers to collaboratively guide the training and learning of the student network, ultimately enhancing the robustness of the student network. In this work, a novel teacher network is developed to enhance the detection performance of the proposed method specifically for small targets. Furthermore, an attention mechanism is incorporated into the student network to enable it to focus on more critical information, thereby improving overall detection efficacy. Experimental results obtained from a measured SAR target detection dataset convincingly validate the benefits of the proposed approach.