Research On The Detection Of Small And Dim Targets In Infrared Images

As key techniques in infrared (IR) homing guidance, target search and tracking, warning and so on, IR small dim target detection and tracking have been regarded as old-line and attractive research topics in the field of IR image processing. As for the real weapon systems, how to make the best of IR target detection techniques to improve the ability of target detection and to obtain the related information about the invading targets, have important significance for improvement of the real weapon systems. The longer the distance of targets, the less the imaging area of targets and the larger the probabilities of targets influenced by backrounds and clutter will be. Therefore, how small dim targets can be reliablely detected under low signal-to-noise ratio(SNR) have become the more realistic and challenging research topics.Aiming at the problem of IR small dim target detection under a complex background, this thesis mainly deals this problem on the basis of the theories of fuzzy reason, high order statistic, transient signal detection and track-before-detect and so on. And the main achievements of this dissertation are as follows:(1) For obtaining better detection performance, based on the classic image difference, a new detection algorithm based on fuzzy reason was proposed, which fused target information from adjacent frames by fuzzy reason theory, and implements 'soft decision'of target following the designing rules. Our algorithm avoids the shortcoming of low detection probability caused by difficult threshold determination and 'hard' decision, and improves the detection performance effectively. In addition, we extends the algorithm to be applied in the infrared dual-band detection and the corresponding algorithm is proposed.(2) Considering the fluctuation of grey-scale values of pixels of the corresponding images, caused by a passing target, as a non-Gaussian transient signal, and combining excellent characteristics of suppressing Gaussian noise, a novel detection algorithm based on third-order cumulant was proposed. The transient signal can be detected by calculating its third-order cumulant. The average value subtraction in the estimation of the third-order cumulant also suppress the background clutter, leading to a great improvement of the signal to clutter ratio (SNR). The experimental results show that the algorithm can effectively suppress heavy infrared background and reliably detect dim targets with SNR larger than 1. Meanwhile, by calculating the Bispectrum likelihood ratio in frequency domain used for target detection, another algorithm was put forward in Bispectrum domain. (3) Duo to the characteristic of the grey-scale value of a pixel in an image occurs fluctuation when a target passes by, the transient signal detection methods was applied in the small dim target detection, and a new detection algorithm based on the Power-Law(P-L) detector was proposed. The higher-order statistics have the advantage of suppressing Gaussian-noise. With the advantage, the method based on Bispectrum slice was proposed, this method can overcome the influence of the background noise and improve the detection performance obviously.(4)Recently, the track-before-detection (TBD) algorithms are effective for the small dim target detection with low SNR. As the tracking part of TBD, the particle filter (PF) is attracting attention. According to the performance descent caused by Monte Carlo random sampling used in the PF and its some improved versions, a new Quasi-Monte-Carlo sampling based Gaussian Particle Filter(QMC-GPF) is proposed. Basing on that, a new TBD algorithm was proposed, which estimates on-line the standard kinematic parameters of the target, including position and velocity, as well as the amplitude of the target. The convergence characteristic of the covariance matrix of the posterior densities propagated in the QMC-GPF is used to determine whether it is the true target. The TBD algorithm, based on the idea of using more regularly distributed point set called low-discrepancy (LD) points to construct the approximation than the random point set associated with MC, would provide the best-possible spread in the sample space and improve the target tracking and detection performance obviously.