SAR automatic target recognition based on convolutional neural networks

We propose a multi-modal multi-discipline strategy appropriate for Automatic Target Recognition (ATR) on Synthetic Aperture Radar (SAR) imagery. Our architecture relies on a pre-trained, in the RGB domain, Convolutional Neural Network that is innovatively applied on SAR imagery, and is combined with multiclass Support Vector Machine classification. The multi-modal aspect of our architecture enforces the generalisation capabilities of our proposal, while the multi-discipline aspect bridges the modality gap. Even though our technique is trained in a single depression angle of 17°, average performance on the MSTAR database over a 10-class target classification problem in 15°, 30° and 45° depression is 97.8%. This multi-target and multi-depression ATR capability has not been reported yet in the MSTAR database literature.