Target Recognition Based On Biological-like Vision Identity

As one of the key aspects of intelligent information processing, target recognition has become animportant component and extremely active branch in computer vision and pattern recognition.The ultimate goal of target recognition is infinitely close to or even beyond in a certain area theability of biological vision, which also stimulates the research interest in the visual perceptionmechanisms of biology. The characteristics of biological vision are mainly as follows: visualattention is a switch, memory is a premise, learning is a necessary condition for intelligenceimprovement, feature association is a material reaction for aptitude, and consciousness is theultimate representation.In this paper, two major aspects of biological vision perception system, signal sparsity and visualattention, are studied and to be used in target classification and recognition system. Here thecritical technologies are focused, including sparse representation for object features, the designof over-complete feature dictionary via learning, target candidate regions detection based onsaliency and multi-feature analysis, and target recognition based on spatial pyramid matchingmodel with thought of “Delamination and Maximization”. The main contents are as follows:(1) To the problem of how to extract essential and pivotal attributes of target, the theory andimplementation process of sparse representation based on effective coding hypothesis are studied.Combined with sparse representation, a discriminative over-complete feature dictionary based onpatch is designed using existed sparse coding and dictionary learning algorithm, which is used tovalidate the effectiveness and feasibility of the sparse representation for signal in the process oftarget identification. The conclusion is that sparse representation has certain discriminativecharacteristics and certain “tolerance” capabilities for the lack of some samples. The reasonableuse of sparse coding can reduce the dependence on the particular characteristics of the target.(2) For visual attention mechanism, the research progress and application of visual saliency areanalyzed. Several classical saliency models simulated by visual attention mechanism, i.e. IT, FT,SR, RC, HC and IS, are introduced expressly. Then, an improved IS algorithm combined withspace similarity analysis is proposed. Experimental results show that the description ability ofnew algorithm for saliency of large target is better than IS model.(3) To address the problems of poor generality and no combination with biological visionperception mechanism for traditional target recognition model, an algorithm of target candidateregions detection based on saliency and multi-feature analysis in a Bayesian framework isproposed, which can possess the discrimination effectively between target and non-target due tothe combination of multi-feature. Depended on the visual saliency analysis, the algorithm has an active selectivity to candidate regions where targets are. In addition, this algorithm has auniversal property in a sense.(4) Several typical target classification and recognition models such as BoW, SPM, HMAX,DPM and Bayesian framework are described and the spatial pyramid matching (SPM) model ismainly studied. Combined with saliency and sparse coding, under the operation of“Delamination and Maximization”, a significant project of target classification and recognition isproposed through analyzing the algorithm of target candidate region detection based on saliencyand multi-feature analysis. In this case, the aim of sparse coding is to describe the features oftarget better and make dictionary more differentiated,“Delamination and Maximization” canensure extracted features have invariance to a certain extent to local spatial transform, andsaliency analysis can remove those candidate areas which have weak capacity when describetarget, thereby the performance of this system is improved correspondingly.In summary, for target classification and recognition, characteristics of visual attention and signalsparsity based on biological vision perception mechanism are simulated. The key issues of howto use sparse coding and saliency analysis to achieve target classification and recognition areexplored. The function of target recognition based on biological-like vision identity is realizedand so theoretical foundation and reference value of this research are provided for furtheranalytical study.In addition, combined with other characteristics of biological vision perception mechanisms,integrating the existing theories, models, algorithms effectively, and to achieve effective targetdetection while ensure efficiency are issues further worth exploring in depth.