Image Edge And Corner Detection Based Upon The Anisotropic Gaussian Kernels

Generally speaking,corners and edges in images represent critical information in describing object features,which play a crucial and irreplaceable role in computer vision and image processing.For many image processing and computer vision applications,corner detection is an essential first step of computation toward a more complicated stream of process,such as object recognition,object tracking,simultaneous localization and mapping(SLAM)and robotic vision,etc.The anisotropic Gaussian directional derivatives(ANDDs)have been proved that are noise robust and provide finer directional intensity variations around a pixel.Thus,it can be used to extract local features of images,such as corners and edges.In this work,we mainly study the applications of anisotropic Gaussian directional derivatives in the field of image processing,such as corner and edge detection.This paper can be summarized as the following two parts:1.Noise robust image edge detection based upon the automatic anisotropic Gaussian kernels.A novel noise robust edge detection method based upon the automatic anisotropic Gaussian kernels(ANGKs)is presented in this paper,which also analyses the problem that Canny edge detector may cause some obvious edges to be missed.First,automatic ANGKs are designed on the basis of the noise suppression,edge resolution and localization precision,it also mediates the conflict between them.Second,thr reason why cross-edge pixels are missing from the Canny edge detector using isotropic Gaussian kernel is analyzed.Third,automatic ANGKs are used to smooth the input image;a revised edge extraction algorithm is used to extract edges from the edge map.Finally,the aggregate test receiver-operating-characteristic(ROC)curves and Pratt's Figure of Merit(FOM)are used to estimate the proposed detector against state-of-the-art edge detection method.The experiment results show that the proposed algorithm can obtain better performance for noise-free and noisy images.2.Multi-scale corner detection based upon the anisotropic Gaussian kernels.A novel edge-based multi-scale corner detector using a statistical method based on the anisotropic Gaussian kernels(ANGKs)is presented,which makes a better performance in corner detection.The noise-robust anisotropic Gaussian kernels can perform better on intensity variations around edge pixels and corners.We utilize the difference of principle directions to build a new corner measure to improve the ability of detection for the local variation and noise on edges.We also analyze the characteristics of the average and standard deviation of the corner responses with different scales and use them to construct the new corner measure.Thus,it is different with the existing multi-scale corner detectors.Furthermore,the new corner measure is shown to have a strong corner response,and it can discriminatethe corner points from the edges correctly.The proposed corner detector is compared with five state-of-the-art edge-based detectors.Three commonly used test images are used to assess the detection capability in different noise level cases.Sixty images with different scenes including some artificial images and many are real world images are used to evaluate their repeatability under various image transformations.The experimental results show that the proposed detector obtains a better overall performance.