Combination of convolutional feature extraction and support vector machines for radar ATR

In this paper a combination of convolutional neural networks and support vector machines for the automatic recognition of ground targets is presented. From the convolutional neural network the feature extraction part is used, but instead of the fully connected multi-layer perceptron in the decision stage a support vector machine is applied. With this combination the generalization capability of the classifier is increased, while the computation time is kept low. The classifier is tested on the public MSTAR database of spotlight SAR data. Results are shown for different kernels as forced decision classifier as well as with rejection class.