A Fast Training Method for SAR Large Scale Samples Based on CNN for Targets Recognition

In recent years, as CNN has made breakthroughs in targets detection and recognition, such method has drawn increasing attention on targets recognition of SAR images. However, when CNN was applied to targets recognition of SAR images, its training efficiency was severely limited by the abundant pixel units of SAR image samples. Compared with CNN commonly used samples, the high resolution SAR images contain more pixel units. If the CNN is directly applied to SAR images, the process of extracting features will have low computational efficiency, which seriously affects the performance of targets recognition. In response to this problem, a method of this paper for preprocessing the input samples is proposed. The experimental results of the real airborne SAR data verify the efficiency of this method.