A Novel Technique to Remove Noise from SAR Images by Equalizing Neighborhood Surrounding Values and Pixel Smoothing

Automatic Target Recognition (ATR) plays a crucial role in military operations to locate actual targets and response immediately. Target recognition is easier in Standard Operating Conditions (SOC) while it becomes very complex in Extended Operating Conditions (EOC) due to environmental noise, large depression variations, partial occlusion and limited training samples. Deep Learning (DL) has a wide applications in Synthetic Aperture Radar (SAR) Imagery. Interpretation of SAR images to recognize actual targets is a major challenge. Researchers from across the world has given different deep learning models for ATR and it is found that most of DL algorithms performs poor in EOC’s. We have increased samples of the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset by adding partially occluded targets, different levels of noise and multiple objects in samples of the MSTAR. This paper presents an Object Localization & Deep Learning-based model for ATR. Our model contains a noise removal phase using spatial filtering and feature extraction and also, classification using enhanced VGG-19. The performance of the proposed model is satisfactory.