| The final target of computer vision research is to make the computer capable of adapting itself to the ambient condition by observing and understanding the external environment with its visual sense. At present, a universal computer vision system, which can identify the objectives in the random background and can independently accomplish various works, could not be developed. But it is possible to devise a special visual information processing model that can automatically completes a specific work. This paper makes a further research in visual information processing model based on existing research conclusions. The main content is as follows:(1)This paper analyses the physiological characteristics of visual system and introduces some of them that is related to the paper and expatiates models of them.(2)Because dual-tree complex wavelet transform is characteristic of multi-scale, multi-orientation analysis, low redundancy and approximately shift invariance, an improved visual information processing model by DTCWT is proposed, which is based on multi-scale and multi-orientation analysis of visual simple cells. And an improved model based on steerable pyramid filters is proposed because of its multi-scale, multi-orientation analysis, shift- and rotation-invariance.(3)To improve performance of target recognition at different viewing angles, a tree layer BP neural network is in use here. The weight is revised during training. At last the performance in target recognition at different viewing angles can be got by testing.(4)An algorithm integrating attention mechanism into the model of visual information processing is proposed. Input image can be bounded by weightbased attention extent. Then our focus is on some points or some area, so the algorithm decreases count and improves recognition rate because of low sensitivity to complex background.(5)The experiments and analysis of improved models of visual information processing are carried on Dcfaces and ORL database. |
PDF Full Text Request