Robust Image Feature Detection Methods Based On Anisotropic Directional Derivatives

For various applications in the fields of computer vision,pattern recognition and computer graphics,the detection and recognition of elementary image feature in images,such as corner and edge,have been an indispensable section or step in the algorithm.For example,the local feature localization based on the corner detection is widely used in the application of object classification and recognition.Besides this,the global descriptor HOG and image semantic segmentation are constructed on precise edge magnitude extraction,meanwhile,the large scale sketch-based image retrieval also needs the image binary edge map.A plenty of theoretic feature models and detection algorithms have been proposed and these primarily satisfy the requirement of low-level feature for most application.In spite of this,there still exist many newly produced problems to be urgently solved,such as the detection accuracy and the noise robustness to different types of noises of the algorithms.Based on the deep analysis on the characteristic of the elementary image feature,the main research of the dissertation is to use multi-scale,multi-direction anisotropic Gaussian directional derivative(ANDD)of images and anisotropic morphological directional derivative(AMDD)to precisely describe the feature and ultimately produces the high-resolution and strong-noiserobustness corner and edge detection.The high-resolution feature detection will help the construction of sophisticated feature descriptor(such as Shape Context),the strong-noise-robustness feature detection will benefit the precise feature descriptor of low-SNR image in the application of video security and protection and has potential valuable study.Based on the ANDD and AMDD representation of the image,the dissertation mainly makes the research on the detection algorithm of corner and edge feature in grayscale and color images and meanwhile tests verify the better overall performance of the proposed algorithms,such as detection accuracy and noise robustness.The main content of the dissertation can be summarized as follow:1.The description capability and noise robustness of ANDD filters for different kinds of edges are analyzed.In the experiments,we discover that the traditional differential operator has some defects in extracting the edge strength map(ESM)of images.On one hand,the sample structure of the operator makes the algorithm sensitive to noise.On the other hand,on describing the complex edge,the result is prone to be interfered by adjacent edge and degrade the edge location accuracy.Based on the study of differential auto-correlation matrix and analyzing the internal relationship between the matrix and edge category.We construct a new metric to measure the edge category attribute and it is used to adaptively adjust the shape of ANDD filter and then achieve the purpose of choosing proper ANDD filter to describe the edge according to the edge category to improve the edge detection accuracy.Meanwhile,ANDD filter shows better noise robustness compared with traditional differential operator.2.To alleviate the conflict between high edge resolution and noise robustness in the color Canny detector,a new color edge detector is proposed by fusing gradient matrix and ANDD matrix.Due to the isotropic Gaussian scale,color ESM by color Canny can't meet the high edge resolution and strong noise robustness at the same time.Fortunately,ANDD filter with proper parameters can satisfy the two aspects except that color ESM by ANDD filter often produces radiated strong edge responses around the corner to lead to false edges.To solve the problem,we give a new insight of color Canny and ANDD-based color ESM in terms of matrix singular value decomposition(SVD)and propose a color ESM fusion strategy based on SVD.The experimental result on real images shows that ESM-fusion-based color edge detection algorithm is superior to the compared algorithms on the edge detection accuracy and noise robustness.3.The noise robustness and detection accuracy of differential-based and morphological edge detection algorithms with respect to different kinds of noise are discussed.The traditional algorithm via differential operators can suppress the Gaussian noise well in images while morphological algorithm robust to salt-and-pepper noise.But,the differential and morphological detector will detect false edges when Gaussian and salt-and-pepper noise both pollute the images.Moreover,these two detectors are bothsensitive to random-valued impulsive noise.To solve the problem,we combine the good characteristic of differential and morphological operator by integrating the ANDD bi-window configuration with weighted median filter(WMF)and propose a new anisotropic morphological directional derivative(AMDD)filter.Embedding AMDD in the routine of Canny detector derives a new edge detector robust to Gaussian?salt-and-pepper and random-valued impulsive noise at the same time.4.The characteristic and defects of state-of-art contour-based and intensity-based corner detection algorithms are deeply analyzed.Based on the divergence between anisotropic differentials of the pixels adjacent to corner and the property of multi-scale anisotropic Gaussian directional derivatives,we design and develop three robust corner detector respectively against to image noise,interference from local image structure and image scaling transform.These detectors are:(1)A corner detection algorithm using multi-scale directional differential ratio(MDDR)of image is proposed to solve the problem that the interference between adjacent image structures may lead to the increase of the false detection ratio.Isotropic Gaussian Differential(IGD)and ANDD filter have different performance for edge and corner pixels.IGD is robust to interference from adjacent image feature but has poor descriptive power for corner feature,versus for ANDD filter.To solve the contradiction,based on the edge detection and contour extraction,the ANDD,whose direction is exactly along the edge direction,and IGD with norm direction are combined to construct ratio-based new corner measure for each pixel on contour.Meanwhile,the fusion of multi-scale corner measures improves the scale robustness of the algorithm.(2)The characteristic and differences of the differential mode at corner and its adjacent pixels are researched.The differential mode at corner has strong anisotropy while good consistence at line-type edge pixel and the differential modes at two sides of corner.Based on this fact,a corner detection algorithm using consistency of local directional differential vectors of pixels is proposed.For each pixel on edge contour,using the consistency of ANDD differential vector of its bilateral pixels constructs a new corner measure.Moreover,the power transform improves the corner distinction,and the corner measure is normalized by the contour length to remove the false corner on arc.(3)Corner detection algorithm via similarity of local multi-scale differential mode is proposed.The local contour around each pixel on contour is divided into left and right support region and then the average differential modes of two support regions and whole region are extracted respectively.The corner measure is constructed from the similarity between the mode averages.The mode average of whole local region helps to find the T-type corner which is difficult to detect using contour curvature-based detector.The fusion of tri-scale corner measures improves the scale robustness of the algorithm.The experimental result on real images shows that proposed detector is superior to the compared detector on the corner detection accuracy.