An Overcomplete Visual Computing Model Oriented Unstructured Scene Object Localization And Its Application

The demand of video and image information experiences a continuous growth in social, economical and security areas, such as Unmanned Vehicle, Video Surveillance and object localization, followed by characteristics like complex objects, scaled media data and diversity requirements. The information can not only be perceived and understood directly, but also processed by computers. In the basic principles of information processing, calculation models and methods, the information processing mechanism of animals based on the visual cortex has the essential difference with the traditional artificial neural network computer vision methods. The processing power and efficiency for complex unstructured scene of the current computer vision can not compare with that of the visual system. Researching and modeling visual mechanism and its computational model is on the verge of a breakthrough for interpreting and verifying brain information processing method, exploring new calculation methods and functional architecture and solving one kind of application problems for complex scenes object localization. Combining visual cognitive and other related science, exploring the actual computational model that is more in line with the biological mechanism and improving the existing computer processing mode represents the main trend of object localization technology research and its development.Therefore, in this paper, focused on extraction, expression and computation concerning visual perception features, based on the assumption of effective coding and visual information sparse coding model, we took the statistical characteristics of natural mages as key point to study sparse overcomplete representation method of analog complex visual information processing mode. Then we established and improved sparse overcomplete visual computing model of simulating complex visual information processing mechanism of the primary visual cortex. And we used this model to improve the traditional method and resolve the problems of capacity and efficiency of object localization of unstructured scene that with many uncertainty factors and unpredictable states. However, the overcomplete representation increases the combinatorial search difficulty of sparse decomposition and changes the symmetry between input space and code space. Thus it makes the model solution and calculation method complicated. Therefore, in order to use the model solves the problem of extracting effective statistical features of natural images, further designed the key algorithm of visual computing model. Then, designed object localization algorithms for unstructured scene based on the above model and algorithm. The experimental results demonstrate the correctness and validity of the above model and algorithms. The main research results as follows:(1) Based on the assumptions that the visual system is result of adapting to the natural environment, we took the statistical characteristics of natural images as point, studied currently animal visual physiological experimental results and information processing mechanisms which are related to the extraction of image statistical properties. And based on this, we summarized and simulated the calculation methods, the optimization criterion and the optimization algorithm of simulating this information processing mechanisms.(2) On the basis of the above studies, combined with the sparse coding mechanism, overcomplete representation mechanism and receptive field mechanism of V1area, established the visual computing model based on the sparse overcomplete. First of all, overcomplete set of receptive fields of simple cells was learned from the natural image sequences. Then the effective characteristics of natural images were expressed. At last, effective information of nature image was extracted.(3) Based on the computable model, an object localization algorithm based on combination of neuron response and dynamic threshold method was further proposed for the object localization problem of unstructured scenes. And multi-class error problem existing in complex unstructured scene was solved by using the error removal method based on the guidance of the object. The multi-class experiments verified the correctness and validity of the algorithm. The experimental results showed that this algorithm can improve the accuracy and the instantaneity of object localization under complex unstructured static and dynamic scene using a small sample.