Influence of MSTAR Azimuth Granularity on CNN Recognition Performance

Abstract

At present, the acquisition cost of synthetic aperture radar (SAR) images is much higher than that of optical images. The smaller the azimuth granularity of SAR target images, the higher the cost of data acquisition. Therefore, on the premise of satisfying the recognition accuracy of the convolutional neural network (CNN), increasing the azimuth granularity of SAR images training set can effectively reduce the cost of data acquisition and improve the effective utilization of data. In this paper, the data of azimuth uniform down sampling in different degrees are used as the training set. The origin images data is the three categories of targets in the Moving and Stationary Target Acquisition and Recognition (MSTAR) database. The target average recognition rate of different training sets is obtained by the CNN model. Finally, we analyze the relationship between the azimuth granularity and the average recognition rate, summarize the influence of MSTAR images data azimuth granularity on the recognition performance of CNN. Based on the recognition of SAR images based on CNN, this paper provides experimental and technical support for azimuth granularity for the future construction of target SAR images knowledge base.