Research On Image Contour Detection Technique Based On Visual Perception Mechanism

Contour information is the crucial part of image after dimension-reduction,which is significant to image applications in subsequent procedure,such as precise and efficient object recognition or image understanding.Traditional contour detection methods that describe the directivity and continuity of contour with mathematical models are usually capable to eliminate unwanted edge produced by background texture and discriminate major contour with non-contour edge.However,in contour detection task,traditional methods always ignore the transmitting and processing procedure of information stream based on visual perception mechanism,which is more like a kind of abstract mathematical description.In this paper,we paid attention to the mechanism of visual perception system in contour detection or extraction,and focused on spikes coding of neural networks and the response characteristics to specific visual stimulus over different levels in visual path,then we attempted to involve concrete applications of image process,such as image contour detection.Firstly,a model named orientative correlation of multiple receptive fields in visual path was constructed,which interpreted the formation mechanism of orientation selectivity and the function of different levels of visual path when carrying on contour detection;secondary,we proposed a method to accomplish the color opponent and population coding of neurons,and applied the method to contour detection for color images;finally,the double-level neural networks with spikes regulative coding was introduced to enhance the modulatory and selective effect for visual information of high-level visual cortex,which could efficiently get rid of the texture over background.The main tasks and research results are listed as follows:(1)The contour detection method called orientative correlation of multiple receptive fields in visual path was proposed in this paper.Then the orientation selectivity of primary cortex was modeled on the characteristics of receptive field at each level in visual path and the transformational characteristics between adjacent levels.The method based on the pre-coding effect of neural network on retina,the adjustment mechanism for non-classical receptive field of LGN neurons,and the orientation selectivity of primary cortex,was utilized to contour detection.Images in RuG library were taken for processing,and the mean value of measure P was 0.43,indicating that method in this paper has favorable contour detection ability.(2)Neurons over sub-cortex usually have the single color-opponent receptive field which can dynamically respond to visual stimulus such as color boundary and lightness boundary;and double color-opponent neurons which are orientation-sensitive on primary cortex can connect with each other by dynamic synapse to code contour information in a population way.In this paper,the contour detection method based on color-opponent and population coding of neurons was proposed to detect images in BSDS500 library,and the mean value of measure P of validation set was 0.55,which indicated the new method is much closer to visual perception system.(3)As the contour detection method based on color-opponent and population coding cannot significantly distinguish the major contour from texture edge,a new contour detection method based on spikes regulative coding mechanism of double-level neural networks was proposed.According to the feedback-regulative principle and selective attention effect of high-level cortex,another neural network of high-level cortex was constructed to capture attention map based on double-opponent neural network at primary cortex.Feeding the attention map back to the response of primary cortex,the final contour was obtained.The mean value of measure P of validation set in BSDS500 was 0.58,indicating this method can distinguish the major contour from texture edge.