Faster All Convolutional Networks for SAR Image Classification

Convolutional neural networks (CNN) has made great progress in the field of object classification in the past few decades. The most frequently used CNN structure is: convolutional layers + maximum pooling layers+ fully connected layers. In the classic structure, convolutional layers and fully connected layers are used for feature map extraction and classification respectively. In the maximum pooling layers, down-sampling is used to reduce the number of calculations of the whole structure. In order to reduce the loss of information caused by down-sampling, an improved version of CNN called an all convolutional networks (ACN), replaces maximum pooling layers with convolutional layers. In this paper, a faster ACN is proposed for SAR image classification. Compared to traditional ACNs, our faster ACN is built by inserting convolutional layers with descending stride, which is helpful for improving efficiency. Finally, experimental verification is performed on the MSTAR data set. Compared with traditional ACNs, our faster ACN can reduce the training time, while maintaining the same level of accuracy.