Classification of Ships Using Real and Simulated Data in a Convolutional Neural Network

Abstract

Convolutional neural networks (CNNs) have recently been applied successfully in large scale image classification competitions for photographs found on the Internet. As our brains are able to recognize objects in the images, there must be some regularities in the data that a neural network can utilize. These regularities are difficult to find an explicit set of rules for. However, by using a CNN and the backpropagation algorithm for learning, the neural network can learn to pick up on the features in the images that are characteristic for each class. Also, data regularities that are not visually obvious to us can be learned. CNNs are particularly useful for classifying data containing some spatial structure, like photographs and speech. In this paper, the technique is tested on SAR images of ships in harbour. The tests indicate that CNNs are promising methods for discriminating between targets in SAR images. However, the false alarm rate is quite high when introducing confusers in the tests. A big challenge in the development of target classification algorithms, especially in the case of SAR, is the lack of real data. This paper also describes tests using simulated SAR images of the same target classes as the real data in order to fill this data gap. The simulated images are made with the MOCEM software (developed by DGA), based on CAD models of the targets. The tests performed here indicate that simulated data can indeed be helpful in training a convolutional neural network to classify real SAR images.