Image Representation Model Based On Primary Visual Mechanism

The human visual system is an important component of the sensory perceptual sys-tem. It contributes to human visual perception, and perfectly enables the abilities of receiving, processing, analyzing and understanding of outside visual information. Sim-ulating the information processing mechanism of visual system and building machine intelligence close to the human intelligence is the key subject of computer vision, pattern recognition and machine learning. A generally-accepted view is that the human visual system has developed gradually with adapting to outside visual information during evo-lution process, and the neural mechanisms of visual system are closely related to statistic structures of visual signals, e.g. natural images. Thus, exploring statistics properties of natural images is instrumental in studying on computation model of visual system.Based on the statistics properties of natural images, this thesis provides theories and computation models motivated by neural mechanisms of primary visual cortex (VI) of visual system, and applies them on solving pattern recognition problem. This thesis consists of two parts, with the first one focusing on theoretical study. Inspired by "sparse-ness", "correlation" and "overcompleteness" of cells in V1, image representation models correspond with the visual mechanisms of V1 are proposed by using natural images. The second part tries to improve application abilities. By simulating and implementing infor-mation processing of the primary visual system, an application model with high practical value is obtained, which proves effectiveness of the theoretical models.The main contributions of this dissertation are as follows:·As traditional linear models are unable to represent complex topographic struc-tures of V1, a pairwise cumulant-based model and a corresponding fast learning algorithm for topographic representations of nature images are proposed. Through exploiting the "sparseness" and "correlation" of cells in V1, the "pairwise cumulan-t" is defined to model binary relations between two adjacent complex cells. Based on independent component analysis (ICA), a two-layer nonlinear model, named P-CICA, for natural images to simulate the topographic structures of V1 is obtained. For the purpose of training this model efficiently, PCICA-F algorithm combining fixed point and Newton iteration method is derived. The local convergence analysis proves that this algorithm is cubic convergence, which overcomes at most quadratic convergence of similar algorithms.·To overcome the limitation of PCICA due to the completeness of ICA model and show the overcompleteness of cells in V1, a quasi-orthogonal estimation based model and a corresponding fast learning algorithm for overcomplete and topo-graphic representations of nature image are proposed. To avoid complexities and inefficiencies of traditional overcompleteness methods, the quasi-orthogonal hy-pothesis of random vectors in high-dimensional space is introduced. According to characteristics of feature extraction task, parts of restrictions of PCICA are relaxed. On this basis, OPCICA model for overcomplete and topographic representations of nature images is obtained. Besides, combining PCICA-F with a quasi-orthogonal iteration method, OPCICA-F algorithm for fast training OPCICA model is derived. This algorithm not only inherits advantages of PCICA-F, but can learn overcom-plete and topographic filters from nature images, which is in accordance with "s-parseness", "correlation" and "overcompleteness" of cells in V1.·Inspired by hierarchical information processing of complex cells, a feature extrac-tion model is proposed to extract features that are invariant to minor variations of input data. First, OPCICA-F algorithm is used to learn filters which demonstrate a clear topography and resemble receptive fields of complex cells. Then the ob-tained invariant complex cell descriptors (ICC), consisting of these learnt filters, are employed to extract the local invariant features from object images. After sim-ple dimensionality reductions of the features, supervised classifiers are trained for recognition. Experiments on the Caltech-101 and MNIST databases show that ICC descriptors, similar to complex cells, can extract invariant features, which over-come linear models' disadvantage of being sensitive to small changes of input. Besides, unlike sparse coding based methods which need to perform some iteration algorithms in feature extraction, our model is not only efficient in feature extraction, but yields comparable classification results.