The Research On Image Segmentation Based On Anisotropic Gaussian Kernel

Image segmentation is based on certain criterion to segment the image into disjoint regions which have specific properties.The traditional methods of image segmentation include threshold-based method,region-based method,and edge-based segmentation method.In recent years,with the development of new theories and new technologies,there are many new image segmentation algorithms have been proposed.The main purposes are concluded as three aspects: first is to suppress noise and improve the accuracy of segmentation results;second is to reduce the computational complexity;third is to make the algorithm more general applicability.The watershed-based segmentation algorithm is widely used in the field of image segmentation,for its simple principle,detecting the edges smoothly and closely,and has a high response to weak edges,etc.However,watershed-based segmentation algorithm is sensitive to noise which can lead to over-segmentation easily;thus,the pretreatment and subsequent processing is necessary.Through the comparison of anisotropic Gaussian kernel and isotropic Gaussian kernel in edge extraction and noise suppression,we obtain that anisotropic Gaussian kernel has better effect on edge extraction and noise suppression.Integrated with the watershed and region merging method,the proposed method improves the segmentation accuracy and efficiency,so as to improve the image segmentation technique.In summary,the paper has following contributions:(1)Certificate that anisotropic Gaussian filter has more edge direction selectivity than isotropic Gaussian filter.(2)Anisotropic Gaussian filter has many directions which has good characterization of edge gradient magnitude and direction.Using the gradient information to extract edge map can solve the edge fracture and edge shift effectively.(3)Anisotropic Gaussian direction derivative filters can suppress noise effectively,which alleviates the over-segmentation problem.(4)Under the region merging cost,region adjacency graph and nearest neighbor graph can accelerate region merging,which effectively reduces the algorithm's time.Finally,the proposed method based on anisotropic Gaussian kernel is compared with Canny,region growing and LOG segmentation algorithms by selecting multiple images,analyzing comparative data from the edge quality evaluation,regional coverage and algorithm complexity.The results show that the proposed method based on anisotropic Gaussian kernel is effectiveness and superiority.