Targets recognition in radar images presents an essential task for monitoring and surveillance of sensitive areas such as military zones. The fundamental problem in radar imaging is related to the recognition of objects in radar images, that its needs a whole chain of treatment. To classify radar images a feature extraction method is used to detect an appropriate subspace in the original feature space, which is based on transformation of the original feature. This subspace should be big enough to maintain minimal loss of information and small enough to minimize the complexity of classifier. Since the feature extractor is difficult to build in manual mode and needs to be redesigned for each application, a Deep Learning in automatic mode is used with a training process subdivided into several modules. In this paper, we lay out an approach to classify Synthetic aperture radar (SAR) and Inverse Synthetic aperture radar (ISAR) images using Deep learning techniques. At first, in order to evaluate the effect of convolution layers and the number of hidden layers of the perceptron we thought of implementing 4 configurations of CNN (Convolutional neural network). In the second time, we use the CAE (Convolutional auto-encoder) to learn the optimal filters that minimize the reconstruction error, after we use these filters to feed the CNN retained and evaluate the effect on performance's system.