Development of automatic target Classification algorithms for sar Imagery

Abstract

Automatic target classification in a Synthetic Aperture Radar image (SAR) plays a vital part in military applications and the exploration on it has been yet emerging. SAR is an active remote sensing system that has the ability to acquire images in the dark and in all weather conditions, thus providing larger dynamic range than optical images. An Automatic Target Recognition (ATR) algorithm extracts meaningful set of features of a target SAR imagery followed by a classifier that performs the target class assignment. In general, an entire SAR-ATR process has six phases: Preprocessing, detection, discrimination, classification, recognition, and identification. But a typical SAR-ATR system fulfils the first four phases. This research puts forward methods for processing Synthetic Aperture Radar (SAR) images that despeckle, detect, classify and recognize targets. The detection of reflected microwaves emitted by air borne or space borne SAR system, results in images dgraded by speckle noise. Speckle noise reduction is necessary in ATR process to precisely detect, classify and recognize targets. In this research work, as a preprocessing step, three despeckling approaches have been proposed to reduce speckle noise in SAR images. In the first method, multiscale directional curvelet transform is applied to various circularly shifted noisy images to remove speckle noise. Though Speckle noise gets suppressed by this technique, the edge preservation remains an issue. Alternatively, in order to preserve fine details on an image, a despeckling in the Undecimated Wavelet Transform (UDWT) domain has been proposed, where the transformed coefficients are subjected to interscale edge investigation. It is found that, this improves edge preservation but the noise reduction is not satisfactory. Always there is a tradeoff between the speckle noise reduction and edge preservation. This warrants the development of an approach where there are two contrary issues handled by modelling it as a multi objective paradigm while filtering in frequency domain. newline