SAR图像舰船检测的神经网络关联剪枝方法

Neural Network Associative Pruning for Ship Detection in SAR Images

合成孔径雷达(Synthetic Aperture Radar,SAR)由于其具备全天时、全天候的工作特点,使其在海洋环境监测、海洋资源调查和海洋防灾减灾等领域得到了广泛的应用。其中,基于SAR图像的舰船目标检测是SAR图像处理中的重要部分,其在军用和民用领域均具有重要的意义。本文针对基于深度学习实现的SAR图像目标检测算法参数计算量大、内存占用率高的问题提出了关联剪枝方法。该方法通过对网络进行改进,将相关联的卷积同时进行剪枝,并在训练结束后统一映射到低维度上以实现剪枝操作。通过在SSDD、SAR-Ship-Dataset和HRSID上进行实验,可以在保证平均精度(AP_(50))下降小于2%的前提下,针对FCOS网络实现70%以上的剪枝率,验证了所提方法的有效性。

The all-day and all-weather working characteristics of synthetic aperture radar(SAR)have made it widely used in the fields of marine environment monitoring,marine resources survey and marine disaster prevention and mitigation.Ship target detection based on SAR images is a critical element of SAR image processing,which is of great significance in both military and civilian fields.In this paper,a deep learning based associative pruning method is proposed to solve the problem of large parameter computation and high memory consumption of SAR image target detection algorithm.By improving the network,the method prunes the associated convolutions simultaneously,and maps them to the lower dimension after the training to realize the pruning operation.Through experiments on SSDD,SAR-Ship-Dataset and HRSID,it is possible to achieve a pruning rate to more than 70%for FCOS networks under the premise of ensuring that the average precision(AP50)decreases by less than 2%,which verifies the effectiveness of the proposed method.