Research On Image Contour Detection Based On Visual Perception

Contour is the segmentation boundary between the target and background.On the one hand,it can simplify and fully describe the geometric appearance of the recognition target,which is of great significance for improving the efficiency and accuracy of subsequent image analysis and understanding.On the other hand,contour sensing is the key ability of biological visual system in interacting with the external environment,which is a good point for the understanding and application of visual mechanism.Therefore,contour detection is one of the hot topics in the field of visual perception.Traditional contour detection methods are mainly based on mathematical models such as gradient operation,they just simply take advantage of the changes in the gray or color of the contour and the background,and do not take into account the inherent mechanism of biological vision in target recognition.Therefore,the recognition of the correct target will become infeasible for the case of serious background texture interference,at the same time,the performance of the contour detection will drop sharply.With the improvement of neuroscience experiments and computing power,some methods based on visual mechanism have been proposed,but most of them like a black box simulation of neural network nature.They are more concerned with the matching of current incentives and responses,and ignore the mechanism of the transmission and processing of visual information flow in the visual path.Therefore,this dissertation focuses on the visual perception mechanism and its application,such as visual path hierarchy,receptive field characteristics and neural coding.Firstly,the receptive field characteristics of primary visual cortex cells were simulated,and the high and low frequency components of visual information flow were respectively taken redundant and sparse coding.Considering the inhibitory and feed-forward effects of neuronal pulse transmission,the differences between the classical and non-classical receptive field response are also fully utilized.Secondly,we consider the azimuthal sensitivity of visual pathway and propose the asymmetric azimuthal sensitivity of neuron receptive field in contour detection.Finally,the influence of different scales of receptive fields on the representation of significant information is studied,and a salient features fusion mechanism based on receptive field and multi-scale is constructed.The scale change is used to realize the classification of the whole and detail of the contour.The main research work and achievements of this paper are as follows:(1)Based on the visual information transmission and processing mechanism in the visual pathway,we propose a new method of contour detection based on the fusion of frequency domain and feedforward mechanism.The incoming visual information is decomposed by NSCT infrequency domain,the high frequency signal is enhanced by redundant coding,the low frequency signal is extracted by sparse coding to extract the effective contour information,and then the above contour information is fused.Gaussian derivative is used to simulate the receptive field of primary visual cortex and introducing the difference characteristics of classical and non-classical receptive field response,the inhibition and feedforward regulation were realized to extract the contour of the image.The dataset Ru G40 is selected as the test object to calculate the comprehensive evaluation index P.The results show that the new method not only can effectively extract the contour,but also has a good inhibitory effect on background texture noise.(2)Based on the multi-orientation sensitive features of visual perception,we propose a method of image contour detection based on multi azimuth selection mechanism of neuron receptive field.Firstly,the image is sparsely coded according to the response characteristics of neurons in the primary visual cortex,and the orientation selection mechanism of receptive fields is improved to construct the response boundary with the proportion of directional weights.At the same time,the receptive field characteristics of the primary visual cortex are simulated based on the nonSymmetric orientation of the sensitive receptive field characteristics of the image receptive field size adaptive selection and effective azimuth response filtering process to obtain the effective contour of the image.After calculating the comprehensive evaluation index P,the result shows that the main body contour obtained by the new method is more complete,the background texture noise is effectively suppressed,and the contour highlighting effect is better.(3)Considering that the single sensorial scale cannot fully reflect the whole and detail features of the contour,we propose a contour detection method based on saliency information multi-scale fusion.Firstly,based on the asymmetric azimuth sensitivity of image receptive fields,we explore the significant information characterization characteristics of sensitive features under different receptive field sizes,constructs a fusion mechanism of significant information,then performs the aforementioned NSCT frequency domain decomposition,and the boundary of the primary cortical boundary and the difference of receptive field were introduced under the improved mechanism to complete the detection of the image contour.The experimental results show that the new method obtains a cleaner outline image,a better suppression for background texture noise and a higher integrity of the main contour.