Adaptive algorithms for change detection in image sequence

Abstract

Detection of small changes/targets in a pair of images in a low signal to clutter plus noise ratio (SCNR) is a problem of great significance in image sequence analysis. The importance of the problem arises in applications such as remote sensing for monitoring growth patterns of urban areas, diagnosis of disease from medical images, visual industry inspection and nondestructive testing for subsurface flaws detection in industrial materials, land resource management, city planning, traffic monitoring, optical and infrared detection from radar images, and weather prediction from satellite and radar images. In this dissertation we present two adaptive algorithms for the detection of small changes/targets (of the order of one pixel) in a pair of images in a low signal to clutter plus noise ratio (SCNR) environment (of the order of $-$14.5 dB). They both have the ability to track the nonstationary image signals (changes/targets and clutter plus noise) and suppress the clutter plus noise background. Both detectors are based on time varying autoregressive models to model image background and on correlation canceling concept. The first one uses an order recursive least squares (ORLS) lattice filter, while the second one is based on a normalized version of the two dimensional least mean square (TDLMS) algorithm. Analytical expression for the improvement factor of the suggested change detectors is presented. Also, the influence of the order of the detectors and of the algorithm parameters of the TDLMS on their detection performances is studied. In addition, the effect of the local mean of the processed images on the optimal estimate of the detector parameters is deduced. The influence of the crosscorrelation between the pair of images on the performance of the detectors as well as the analysis of the computational loads for the processors are studied. The performance of the two algorithms are evaluated by using an optical satellite image, as a clutter background, with computer generated target and noise added to it.