Research On Edge Detection Of Color Image Base On Structure Tensor

Edge detection is one of fundamental tasks in computer vision and image processing, andthe detected edges not only represent some important features in a image, but also providepriori knowledge for the subsequent image processing. It is a challenging task to detect theedges in gray-level images especially color images. The studying of the methods for colorimage edge detection and the improved quality of edge detection, affect the quality of thesubsequent image processing, analysis and understanding.Edge reflects the discontinuity and difference in image information which can bedescribed by gradient. Therefore, the computation of gradient for each pixel is a key issuein edge detection process. For color images, due to processing the coupling between colorchannels, the method based on structure tensor to calculate the gradient of a color image isbetter than the general method which computes gradient of the given image by processingthree channels separately. However, in general, the gradient magnitude calculated bystructure tensor is not rotationally invariant, that is, the gradient magnitudes calculated bystructure tensor at same edge point are not consistent before and after rotation, which isnot isotropic. In order to resolve this problem, this paper designs a novel method tocomputation of gradient of a color image, which is based on structure tensor and isrotationally invariant. The method can be described as follows. First, the gradientorientation of each pixel in a color image is calculated by structure tensor approach. Then,the Gaussian filter is used to smooth the image along the vertical director of the gradientwhich can reduce the influence of noise, and then the gradient magnitude is obtained byGaussian subtraction method along the direction of gradient. After that, the edges aredetected by some operations like Canny method.We compared our method with the state-of-the-art edge detection algorithms includestructure tensor-based method, compass-based method and Canny method, and thequalitative and quantitative results demonstrate that our method outperforms the others.