Research On Object Contour Detection Based On Visual Perception Mechanism

Visual perception is the key content of biological intelligence.Through the understanding of images or videos,people can map the information of the object ive world into the perceptual content in their minds,and target contour detection is an important basis to achieve the above functions.Due to the outstanding performance of biological vision system,it will be helpful for the construction and application of visual computing model by deeply studying its internal working mechanism.In this paper,we simulate the perception and expression process of visual information flow.Based on the analysis and transmission of visual information flow in visual pathway,a contour detection model of visual pathway based on variable scale receptive field is proposed.Considering the difference and correlation of response of hierarchical neurons,a contour detection model with X-Y dual channel and multi-inhibition is proposed.Based on the contour detection of X-Y dual channel and multi suppression,a semi supervised classification method of multi branch feature joint decision is proposed.The main contents and achievements of this paper are as follows:(1)In this paper,a variable scale perceptual model is proposed.The traditional contour detection does not consider the relationship between visual features and receptive field scale,and ignores the difference and correlation of visual information between different visual ch annels in biological vision system,resulting in blurry details of contour detection,insufficient texture suppression and unsatisfactory contour detection results.In this paper,the mechanism of visual information processing based on multi-scale and sparse visual field is proposed,the visual pathway was reconstructed according to the different visual background information.Finally,the validity of the algorithm model is verified by the image of Ru G40 library.(2)A contour detection model based on X-Y dual channel and multi suppression is proposed.Considering the classification of neurons and horizontal cells and the independent and parallel characteristics of visual information transmission channels,a mathematical model of X-Y dual channel information transmission is proposed;in order to adapt to the visual detection task under the distribution of contrast difference,a dynamic contrast adjusting receptive field response model is proposed;a variety of texture suppression methods existing in the visual processing mechanism are simulated,and a mathematical model of X-Y dual channel information transmission is proposed In order to ensure the integrity of contour structure,the input ring suppression and sparse coding are used to realize the texture suppression of X and Y channels respectively;the automatic contour completion function of biological vision system is simulated,and the orientation is subdivided based on the optimal orientation,and the neuron response of fine orientation is captured by using two-dimensional Gaussian derivative,so as to ensure the integrity of contour st ructure.Finally,with rug40 as the experimental library,the binary contour map of the experimental results is compared with the reference image,and compared with the mainstream contour detection methods.The results show that the comprehensive performance of the proposed method is better than the comparison method.(3)On the basis of the target contour detection model of dual channel and multi suppression,a semi supervised classification method based on multi branch feature joint decision is proposed and applied to image classification task.Firstly,the target contour is input into the classification network,which effectively improves the classification efficiency and complexity;in addition,considering the massive demand of the traditional classification network for the labeled data,and the poor network feature representation ability and low feature utilization rate,this paper improves the representation ability of the original contour data through the diverse and complementary multi branch features,and designs multiple branches Finally,the multi branch feature network is trained according to the pseudo label data and labeled data to further improve the feature extraction ability of the network.In this paper,the market-1501 data set is selected for the experiment,and mean average precision(m AP)of 56.8% and rank-1 accuracy of 80.3% are obtained,and good performance is achieved.