Research On Neural Computing Model And Its Applications Based On Visual Perception

Visual perception is one of the hot research problems in computational neuroscience and neural engineering.On the one hand,it helps to reveal the inner mechanism of the amazing ability of the visual system to perceive information.On the other hand it will bring new ideas to the machine vision and other intelligent information processing methods.In this paper,neural networks based on human visual perception mechanism were constructed to explore the neural information processing using the spatial-temporal characteristics of firing spikes,and to try to deal with practical applications in image processing.Firstly,the proposed method designed new neurons receptive field model in the form of synaptic connections.Through the model response to the light and shade stimulus,we explained the important role of the new model in achieving the vision mechanism.Then,a under visual cortex-primary visual cortex hierarchical processing model was constructed considering the multi-stage directional sensitive nature of the visual pathway.We used the results of edge detection in low-contrast images to explain the key role of the orientation sensitive mechanism plays in the contour perception.Finally,sparse coding rules and visual attention mechanism were introduced to build neural network which had the characteristics of spatial-temporal information fusion.With the modulation of boundary response by saliency image,we provided a new way of image processing applications based on visual feedback control.The main tasks and research results are listed as follows:(1)New receptive field models were designed in the form of synaptic connections to implement orientation sensitivity,lateral inhibition and selective attention mechanisms.These models could give full play to their roles in the perception of contrast and spatial variation in visual stimulus.Took the image feature which was integrated with the spatial information of receptive field as input of neuron model to convert visual stimuli into a pulse sequence.And characteristics of visual stimuli were encoded in the exact time the spike fired.Neuron responses indicated that these new models could perceive direction,detect shading edges,and could implement other preliminary visual function.(2)Considering the multi-stage directional sensitive nature of the visual pathway and dynamic characteristics of synaptic connections,a hierarchical image edge detection method was presented.Using the physical structure feature of ganglion cells and LGN neurons receptive field distributing centripetal,a sub-cortex multi-direction sensitive function layer was constructed to get the edge sensitive strength of an image with fuzzy hierarchy and rich details.Then a primary visual cortex function layer was formed with optimized lateral inhibition to remove redundant information.Experimental results showed that this method could completely detect image edge and effectively filter out texture noise.(3)For visual pathway model didn’t have the ability to distinguish between the body contour and texture edges,a new contour detection method base on sparse coding of salient region was proposed by introducing the neural sparse coding rules and visual attention mechanism.For the orientation perception of V1,a sparse neural network with spatial information was instituted to suppress the response of texture edges.The difference characteristics of classical receptive field and the non-classical receptive field was calculated and introduced to the network input,and the network output saliency image.Finally,the V1 contour response was obtained by modulating the feedback of saliency image with its orientation perception.In the contour detection task of natural images,the new method was proved to be can effectively inhibit the response of textured edges,while retaining as much of the real contour,and had superior stability.