Research On Visual Recognition Based Perceptual Computing Model

With the high resolution imaging and the hardware ability of parallel computing,the analysis and comprehension technology based on massive visual data has become one of the research focuses in the pattern recognition and artificial intelligence.To some extent,the tradition visual recognition algorithms consider the principle of the biological perception as well,such as using the massive training sample data to modify the structure and parameters for neutral network dynamically and realizing so-called optimal decision.But they often only use some basic characters of biological perception and simulate it as a black box overall.Actually the abundant visual mechanisms in biological perception system are the bases to realize the visual comprehension and recognition.Therefore,this thesis focuses on the visual information flow processing mechanism in the biological perception system,tries to establish a new perception computing model which faces to the visual recognition,and takes the common face verification and recognition in visual recognition as an example to carry out the study.The contents such as the multi pathway delivering in the visual path for visual information flow,independent recognition mechanism for upright and upside-down faces and multi-scale face recognition are studied in this thesis.The details and results are as follows:(1)A visual information flow dual vision pathways based face verification model is proposed.A convolution network at the bottom is established to simulate the filtering of visual information flow by the biological perception system;the concept of dual vision pathways is proposed to construct an enhanced network of supervisory signal and recognition signal fusion and to realize the enhancement of visual recognition features;finally by comparing the identification photo and multi-faceted photos of the spatial distribution,the joint Bayesian face recognition method is used to identify the similarities and differences of faces.The experimental results show that the design of dual vision pathways can improve the accuracy of face verification and the supervisory signal can restraint the face feature extraction.(2)A new upright and upside-down faces recognition method based on the cooperation mechanism of left and right hemispheres is proposed.To simulate visual information flow transmission and processing in the visual pathway,the bottom neural network and the sensitive texture features and symmetric convolution kernel mechanism are built and the upright and upside-down faces image removal redundancy and preprocessing are achieved;then a concept of pooling neural network layer extracting from local regions is proposed and a multi-local feature fusion network structure is constructed to compress and extract local information.Finally,according to the characteristics of the left and right hemispheres in the high-level visual cortex,a prediction function which integrates global and local information is proposed.The experimental results show that the proposed method can identify both upright and upside-down faces.The results also indicate that although they play a decisive role in recognizing orthographic and inverted faces,the two visual pathways of global and local features are not isolated but complementary.(3)A new multi-scale face recognition method based on multi-level information expression is proposed.First of all,the sensory characteristics of retina and ganglion cells to the underlying texture features are simulated and a multi-layer neural network with convolutional pooling structure cascade is constructed to filter and refine the visual information.Secondly,considering the benefit in the bottom information flow expressing the face character,the concept of multi-scale pooling based on local and global integration is constructed and the fixed-size feature channel and feature map is output.Then,according to the decision mechanism of visual cortex,the feature information of each level is integrated into the fully connected layer and the activation characteristics of neurons are used to determine face identity.The experimental results show that this method can be effectively applied to the recognition of multi-scale face images and the results also indicate that the underlying texture features is helpful in the face features expression.