Contour Detection Model Based On Scale Information Of Visual Receptive Field

Contour detection is the basic task of image segmentation,pattern recognition and so on.It is an important research field of computer vision.Different from the traditional edge detection,contour detection can distinguish the detected edges from the background texture or contour and suppress the background texture and retain the target contour,which is more complex than the edge detection.The human vision system can quickly acquire the target contour in natural scenes,which inspires us to detect contour by imitating human vision system.Physiological studies have found that the response of neurons in the primary visual cortex(V1)is the result of the Classical Receptive Field(CRF)and the non-classical Receptive Field(n CRF),which can be effectively used to eliminate the background texture.Considering the traditional contour detection models use single scale to detect contour,Contour detection models using multi-scale information is proposed in this paper.First,a contour detection model based on parallax information is proposed.For each pixel in the image,when using function filtering,the parameter of receptive field are set according to parallax information,which is modulates by sigmoid function.Different from the traditional edge detection,the scale used for filtering is different for every pixel.The experimental results of NYUD-V2 image database show that the result of multi-scale contour detection is better than the effect of the traditional model and have higher performance index.It is difficult to integrate results under different scales.A multi-scale contour fusion method is proposed.At first,based on the physiological characteristics of the non-Classical Receptive Field,multi-scale contours with surround inhibition are obtained.Based on the minimum scale contour binary map,the current pixels has the optimal direction on both sides are used to determine whether there are pixels on other contour maps.After that,the Gaussian function is used to weight the relevant information and obtain the weight graph of the image.Finally,the weight graph is subjected to non-maximum suppression and binarization to obtain the final results.Experimental results based on Ru G40 database and Berkeley image segmentation database show that comparing with original method,the proposed model can enhance the weak contour information in small scale,and suppresses the background texture in large scale.It improves the detection effect of the original contour detection method and enhances the robustness of original method.