Detection And Classification Of Edges In Color Images

Edge detection is one of the most important and difficult tasks in computer vision.It requires accurate edge detection and classification.However,edge classification is difficult or even impossible in gray-level images.Color images can solve this problem by introducing extra color information.The theoretical analysis and experimental results in this thesis show that the existing color edge detection algorithms easily lose the low-contrast edges and underutilize the color information while the existing color edge classification algorithms are sensitive to illumination color and often can but classify special edges.This thesis is targeted for an appropriate solution to these problems.Edge detection algorithms can be divided into three categories:output fusion methods,multi-dimensional gradient methods and vector methods.The comparison of the three methods indicates that the vector method is the best one.However,all these methods do not make full use of regional information which results the easy losses of low-contrast edges.So,a new color edge detection algorithm based on regional distance measurement is proposed in this thesis.A 3×3 mask in the image is adopted and pixels in the mask are divided into two sets according to the ideal edge model. Then the regional distance of the two pixel sets,which is used to generate edge and direction map,is calculated using vector distance matrix.Finally,the non-maxima suppression is employed to extract the edge points.On the other hand,accuracy of color edge detection does not only rely on detection methods but also on color space.The existing methods are based on RGB color space and neglects color information.Besides,RGB space has other drawbacks: strongly correlated components,poor human perception,non-uniformity,etc.So RGB space is not suitable for the edge detection.HIS is another color space which separates the intensity and color components and benefits for developing color description-based detection algorithms.Thus,edge detection is accomplished in an improved HIS space which removes the saturation normalization by lightness and is more useful for application.Experimental results demonstrate that the proposed method outperforms other recently proposed color edge detection algorithms in RGB and typical HIS spaces for partial test images.Since RGB color space does not utilize color information enough,it can not classify edges efficiently.In this thesis,Dichromatic Reflection Model is used to analyze the invariant properties of classical color spaces.Experimental results show that these color spaces can only detect a particular type of edge and lose the edge discriminative power.Furthermore,another existing differential-based edge classification algorithm is compared with preceding algorithms,which are based on the zeroth-order image structure,and is evidenced to have a better ability to distinguish edges than them.However,no matter the differential-based or zeroth-order-based classification algorithms are sensitive to illumination color.In order to overcome this shortcoming,a novel edge classification algorithm based on color constancy is proposed to corrects images from different illumination to a canonical light source and reduce the dependency of classification on illumination color.Experimental results show that such algorithm can obtain robust results.