Random field modeling and joint detection/estimation filter for multi-sensor image fusion

The problems encountered in automatic target recognition (ATR) are studied in this thesis, especially those problems in the ATR systems using forward looking infrared (FLIR) and laser radar (ladar) sensors. The emphases of this thesis are placed on (i) the development of a novel target segmentation algorithm, (ii) the multi-sensor data fusion for enhanced ATR, and (iii) classifier design.; Although there exist quite a number of segmentation algorithms, none of the algorithm can be used for selective segmentation, which is critical for an efficient ATR system. We propose a novel segmentation algorithm that can perform selective target segmentation. The algorithm is based on the conditional Markov field (CMF) modeling of a 2-D image. The CMF is a composite random field. In addition to the gray level, a random parameter {dollar}theta{dollar} is introduced to incorporate the local characteristics of a 2-D image. A joint detection/estimation filter (JDEF) is developed to estimate the statistical features at the pixel level by adaptive Kalman filtering and extract the possible targets by Bayesian decision theory.; Multisensor data fusion technique is widely used to enhance the performance of the ATR system. However, the fusion of FLIR and ladar data has limited success because the signals received from the different frequency channels have entirely different characteristics. In this thesis, the characteristics of the FLIR and ladar data are carefully explored. A fusion scheme based on the particular features of the FLIR and ladar images is proposed for segmentation. Only the regions of interest are processed which greatly reduces the amount of data need to be processed.; The design of pattern classifiers is investigated in this thesis. The existing classifiers are analyzed, and a fuzzy logic classifier is developed.; One of the main contributions of this thesis is the novel target-oriented segmentation algorithm. The proposed segmentation algorithm, based on inhomogeneous random field modeling, estimates both the mean and variance of the signal at each pixel thus allowing efficient extraction of the man-made objects characterised by relatively homogeneous random field. The main application area of this segmentation algorithm is the ATR system. But it can also be applied to fingerprint processing and medical image processing. Other main contributions include a fusion algorithm for FLIR and ladar data which allows fast and accurate target extraction, and the fuzzy logic classifier which is capable of performing information fusion and imprecision-tolerant classification.