Visual Processing Model And Its Research On Application Based On Non Classical Receptive Field

Edge or contour as the key features in images,which have a great significance for the analysis or understanding of the target object in visual field.Recently,the edge or contour detection methods are mostly based on the jump characteristics of the pixel brightness and other attributes by mathematical description.So from the image space point of view,the jump characteristics usually reflect the change tendency of the pixel properties in the local area.But when the visual system extracts the edge or contour of the image,it always refines the local details of global or large-scale regional features,and the disinhibition region of non-classical receptive field is the basis for achieving the visual mechanism.Therefore,this paper focused on the characteristics and applications of non-classical receptive field.Firstly,the traditional three Gaussian function model was improved,we used non-classical receptive field structure to realize the adaptive weighted response of the input,in order to enhance the edge information of image.Secondly,we improved the traditional homogeneous inhibition model,and used color space information to inhibit the heterogeneous components,then we obtained the edge details,which had the better anti-interference ability.Finally,we simulated the extraction and feedback of visual salience information,and then we combined with the texture homogeneous inhibition model,so as to highlight the main contour.The main tasks and research results are listed as follows:(1)Three visual processing models based on non-classical receptive field mechanism were established.Firstly,an adaptive weighted response model was proposed,which based on the three Gaussian function model of non-classical receptive field,in order to get more complete edge characteristics.Secondly,we improved the traditional homogeneous inhibition model,which based on classical receptive field,the three Gauss function was introduced to replace the DOG function to simulate the non-classical receptive field.Lab color space and slice similarity were used to describe the color and texture similarity of image pixels,in order to obtain background texture and heterogeneous components of images.Finally,a salience information extraction model was proposed.According to the difference between classical and non-classical receptive field,the synaptic inhibition current was generated,and the exponential normalization was used to extract the salience information.(2)A new edge detection method based on non-classical receptive field mechanism was proposed.Firstly,we used the Log-Gabor filter to select the direction of the image information,we adopted izhikevich neural network and sequential coding to simulate the visual channel,so as toobtain the preliminary edge image.Secondly,the edge information was enhanced by the adaptive weighted response model.Finally,the color homogeneity inhibition model was adopted to remove the heterogeneous components,and the final edge image was obtained.The proposed method was evaluated as following,the average area under the ROC curve was 0.850,and the average value of information entropy was 0.395,both of them were superior to the traditional detection methods.(3)An image contour detection method based on visual salience information was proposed.Firstly,the derivative of Gaussian transform was introduced to get the gradient responses of primary visual cortex(V1)in multiple directions.The synaptic excitatory and inhibitory currents were specified based on receptive field mechanism in the LIF-neuron network.Then visual salience information was obtained by coding the spike frequency.And then the responses of neurons in V1 were modulated by feedback to get the image contour response.Finally,the image contour response combined with texture homogeneity suppression model to get the final contour.The pictures used for this experiment were selected from the RuG library.In this paper,the mean value of measure P was 0.50,which had a significantly better performance than the traditional methods including ISO.