A Fuzzy Approach to Region of Interest Coding in JPEG 2000 for Automatic Target Recognition Applications from High-Resolution Satellite Images

Abstract

Automatic target recognition (ATR) from satellite imagery involves detection of foreground (FG) objects from the background (BG). ATR demands higher fidelity, which in turn requires more bitrate, hence a conventional compression, which does not discriminate targets with the background results in poor detection rate. Here we propose a mechanism to achieve lower bitrate without compromising the detection efficiency. By allowing the background to be coded with lower fidelity than the regions-of-interest (ROIs), significant gains can be achieved in terms of compression and hence in storage space and transmission times. One interesting feature of the new JPEG-2000 image coding standard is support of ROI coding using maximum shift (maxshift) method, which allows for arbitrarily shaped ROI image compression without shape coding or explicitly transmitting any shape information to the decoder. We propose a fuzzy C-means clustering approach to generating arbitrary shape mask so as to cluster the images into regions of varying homogeneity. Homogenous ROIs can be coded at a lower bitrate than the high detail regions. This ensures that the target recognition process is not affected by the compression process. A validation benchmark using 'fuzzy feature vectors' is proposed which checks the foreground objects for features like rectangularity, circularity, elongatedness, symmetry, area etc., as compared to its uncompressed equivalent. The validation is done on Standard JPEG, JPEG-2000 with ROI coding and JPEG-2000 without ROI coding at preset bitrates and compared.