Target Recognition From Multispectral Night Vision Images Based On Biological Visual Models And Complex Information Learning

The explorations of new methods for intelligent scene understanding and target perception according to the characteristics of multispectral night vision images have been an inevitable trend in the development of night vision technology. By introducing the biological visual models and complex information learning methods into the target detection and recognition from multispectral night vision images, a series of calculation models and learning methods with the typical characteristics of human visual perception and brain cognitive function are constructed in this paper. The proposed models and algorithms can bring about the robust recognition of night vision targets in natural environments, and the new technologies can provide intelligent understanding and detection from the multi-spectral and large-scale night vision information. There are two problems to be solved for the target recognition from multispectral night vision images based on biological visual models and complex information learning methods. One is the bionic visual modeling applicable to the features of night vision images in multiband, including the studies of denoising, salience analysis and target recognition in night vision images. The other is the complex information learning in multispectral and large-scale night vision image set to implement multispectral target detection and target dimension reduction, classification and recognition in large-scale night vision data. Aiming at these two issues, the main research works in this paper can be divided into the following aspects.(1) Low-light-level (LLL) image enhancement based on a local sparse structure denoising model (LSSD). The classical visual sparse and redundant representations perform well in image denoising. However, they fail to denoise the complex LLL images and tend to lose the structure information in image. In this paper, by embedding the noise invariable features of local structure into the sparse decomposition process, a local structure preserving sparse coding (LSPSc) algorithm and a kernel LSPSc (K-LSPSc) algorithm are proposed to improve the stability of sparse representation for image patches corrupted by heavy noise. Based on these two algorithms, a LSSD model is constructed for noise inhibition and texture details preservation in the nature LLL images, which can increase the signal-to-noise ratio in LLL images effectively.(2) Salient contour extraction from night vision images based on the non-classical receptive field (nCRF) modulation models. Firstly, in allusion to the interference of noise and singular data in LLL iamges, a WKPCA (Weighted Kernel Principal Component Analysis) DH (Degree of Homogeneity) amended nCRF inhibition model is proposed. By evaluating the differences between center-surround comprehensively and accurately, the presented model effectively improves the precision of salient contour extraction from LLL image with complex scenes. Secondly, to solve the problem that there are contour fractures caused by blurred details in night vision images, the inhibition and facilitation effects of nCRF are incorporated to present a compound modulation model. It can strengthen the response of weak contours and connect discontinuous contours to extract the integral salient contours from natural scenes in the LLL and infrared images precisely.(3) Robust night vision target recognition based on a local sparse structure matching model (LSSM). Visual sparse classification performs well in target recognition, but it has high dependence on templates and weak ability of generalization for general night vision targets identification in complex scenarios. Based on the robust local structure preserving sparse coding (LSPSc) and kernel LSPSc (K-LSPSc) algorithms proposed in this papre, a LSSM model is constructed by combining a concept of local matching. Without training for the interested targets with extensive and comprehensive templates, the LSSM model can realize target identification based on a simple template set. Besides, it has steady cognitive ability in recognition of night vision targets with situations of the variations of scene, differences of target shape and occlusions of background.(4) Target detection and recognition from multi-spectral and large-scale night vision images based on the complex information learning methods. Firstly, because the multispectral night vision image set is influenced by the morbidly distributed data, the target detection is inaccurate. A spectral angle matching (SAM) weighted kernel eigenspace separation transform (KEST) algorithm is proposed to suppress the interference of anomalous data and improve the detection rate of multispectral night vision images remarkablely. Secondly, it has problems of outliers interference in the manifold dimension reduction of night vision data. A kernel maximum likelihood (KML) weighted kernel local linear embedding (KLLE) manifold algorithm is designed to detect the outliers and optimize the neighbors selection accurately. The proposed algorithm can enhance the reliability and separability of the low-dimensional representation for night vision data, and can achieve the precise and efficient classification and identification for a large-scale night vision image set.