The Study On Extracting Features From Color Images

Edges and corners are fundamental features of images.Edge detection and corner detection are widely used in many image processing fields,such as image fusion,object recognition and 3D reconstruction.Color images contain more information than gray-scale images,therefore this thesis focuses on the edge detection and corner detection in color images.The first part presents an edge detection algorithm of a color image through multi-pixels anisotropic Gaussian directional derivative filters(MAGF).The MAGF has fully taken into account the fact that the edges have lengths,and they have the properties of elongation,continuity and anisotropy.Meanwhile since the filters attract the advantages of Gaussian filter and mean filter,they could effectively suppress noise and preserve well the features of edges.Firstly,MAGF are constituted by a set of separation filters with K directions in[0,;?).Secondly,the multi-pixels anisotropic Gaussian directional derivative matrix(MADD)is stacked with the convolution of the three components of a color image in the RGB space with MAGF.Thirdly,the 3×K matrix indicates the local gray-scale variations of all three components at each pixel.In order to achieve the edge fusion of three components of a color image,the singular value decomposition(SVD)of the MADD is adopted to obtain the MADD color edge strength map(CESM)and color edge direction map(CEDM).Fourthly,the non-maximal suppression operation is used to eliminate the spurious response of edge detection based on the CESM and CEDM.Finally,the effectiveness of the proposed algorithm is experimentally demonstrated with the qualitative and quantitative(by aggregated ROC curve)comparisons of the proposed color edge detection algorithm with other algorithms both in noiseless and Gaussian noisy cases.In second part,a color image corner detection algorithm is proposed.The corner detection is based on the ratio of the two second to the two largest local maximums(also called peak)of fusion results that obtained from the anisotropic directional derivative(ANDD)representations of three components of color image.We calculate the ratio of the two second to the two largest local maxima by the l2-norm fusion of ANDD representations of three components of a color image.The corner is detected by this ratio and direction difference of the two second and the two largest local maximums.The proposed algorithm is compared with the improved Harris color corner detection algorithm through the statistical analysis of the experiment in several color images.In third part,a color edge detector using the gradient matrix and anisotropic Gaussian directional derivative matrix is adopted in the curvature scale space corner detection algorithm with adaptive threshold and dynamic region of support to design a new color corner detection algorithm.The designed algorithm is directly compared with the improved Harris color corner detection algorithm through several color images.