Synthetic aperture radar(SAR) ship detection plays an important role in ship dispatching, battlefield dynamic tracking and other applications, which all require real-time inference. High spatial resolution of SAR images means a large amount of data and a large computation cost, which makes it difficult to realize real-time inference in hardware with limited resources. Therefore, we propose a new model compression scheme to learn a slim ship detector named Lite-YOLOv4. By modifying network structure to be lightweight and improving polarization-based channel pruning, we generate a compact model. In order to further compress the model, a progressive training-based mixed-precision quantization is proposed to simplify the model bit-representation, reduce its storage requirements and reduce the amount of computation. Finally, the model performance is restored by a self-designed target-guided module with attention mechanism. A lot of experiments were carried out on SSDD, which verify that our method is advanced compared with other CNN-based algorithms. The proposed detector with 0.64MB parameters, achieves 96.6% AP with the calculation cost of 36.11 bitGOPs, in which bitGOPs is 70.9% lower than SOTA, parameter storage is 46.7% lower and AP is 2% higher accuracy.
|