Design of Convolutional Neural Networks for Classification of Ships from ISAR Images

Automatic recognition of ships is crucial and important in military applications for maritime surveillance. Inverse synthetic aperture images of ships generated from maritime radar aid in classification of ships. With the advent of convolutional neural network (CNN) and being successfully used for object recognition in computer vision applications, the use of CNNs for radar images is significant. Visual Geometry Group (VGG16) is a type of convolution network that is used for image recognition and classification. Transfer learning with pre-trained CNN using VGG16 model is used for classification of ships from Inverse Synthetic Aperture Radar (ISAR) images. These models can be further optimised to classify ships with few features because they are pre-trained for optical images with a variety of classes. In this paper, we construct and propose a convolutional neural network model, an optimized VGG16 network model with reduction in number of layers and number of filters for classification of ships from ISAR images. The model is robust and has led to better convergence with an improved classification performance and simplification of the network. The number of layers are drastically reduced and the number of filters are designed using clustering technique. The design of this simplified architecture is discussed in this paper and is compared with VGG16 based model with transfer learning and has shown better performance. The dataset consists of electromagnetically simulated targets at different azimuth and elevation angles. Also, a real time scenario depicting the oscillatory motion was also used to generate ISAR images. The proposed CNN has shown good performance for this scenario as well.