Study And Application Of Image Segmentation Algorithms

Image segmentation is a procedure to separate an image into homogeneous regions that are more meaningful and easier to analyze. Image segmentation is the basis of image processing, and plays a key role in computer vision and object recognition. In recent years, many related papers have been published, which make the accuracy and efficiency of image segmentation improved. However, there are still some problems to be improved. In this paper, on the basis of the investigation and summary of traditional image segmentation methods, we improve some classic image segmentation methods, and the primary works and remarks are as follows:(1) A global spatial similarity based FCM algorithm was proposed to solve the problems that the segmentation results of the traditional FCM based image segmentation algorithm are discrete in the spatial distribution. A global spatial similarity measure and a global intensity similarity measure are proposed and introduced into a novel distance metric to calculate the distance between the pixels in the image and the cluster center, adjust the free parameter to control the proportion of the spatial feature and the intensity feature in the feature space, which enhance the spatial distribution continuity of the segmentation results and improve segmentation accuracy.(2) The image spatial feature is introduced into Fuzzy C mean cluster algorithm, which can impose spatial continuity on the segmentation results. However, the image spatial feature can also to some extent limit the spatial distribution of pixels in each class of the segmentation result. To solve this problem, a similar class merging based FCM algorithm was proposed for image segmentation. Color histogram is used as a descriptor to represent each class segmented by the FCM initial segmentation, Bhattacharyya distance is used to calculate the similarity between any two classes, and a maximal similarity based class merging strategy is used to merge the classes with maximum similarity to obtain the final image segmentation results.(3) To solve the problem that traditional genetic based clustering algorithm and ant colony based clustering algorithm are easy to converge to local optimum, an improved harmony search algorithm was proposed for clustering analysis. The feedback mechanism was introduced into harmony search algorithm, and harmony memory considering rate and bandwidth can be dynamically adjusted by calculating the difference between the best harmony and the worst harmony, which make the proposed algorithm converge to the global optimal solution quickly. A method which can automatically determine the right number of clusters of data samples was proposed.(4) An improved FCM was proposed and used in interactive image segmentation to solve the problem that traditional FCM based image segmentation algorithms can’t achieve accurate image segmentation by their own limitations. User inputs seeds as cluster centers, and image spatial feature, intensity feature and texture feature are introduced into FCM algorithm to calculate the difference between pixels and cluster centers. The final image segmentation results can be obtained by calculating the membership between pixels and seeds.(5) To solve the problem that PSO based Shortest Path algorithms are inefficient and easily converge to local optimum, an improved global harmony search algorithm is proposed to solve shortest path problem. A dynamical genetic mutation probability is proposed and introduced into harmony search algorithm, which can effectively prevent the proposed algorithm from trapping into the local optimum. A dynamical priority-based encoding was used for harmony representation in the proposed algorithm, and a path will be built according to the value of decision variable in the harmony vector. The shortest path will be obtained through updating harmonic memory.(6) A PSO-based live wire interactive image segmentation algorithm was proposed to solve the problems that the objective contour is easily influenced by imaging artifacts and imaged at low computation speed. The gradient amplitude change function between adjacent nodes was introduced into the new cost function to reduce the interference due to imaging artifacts and improve the segmentation accuracy. To improve the implementing efficiency of the proposed algorithm, PSO was applied to finding out the shortest path between any two points in image so as to locate the objective edge.(7) An adaptive random walk image segmentation algorithm was proposed to solve the problems that the description of image information is simple and the objective contour is easily influenced by natural texture. Texture feature is introduced into random walk algorithm to highlight image structural information, which works with image intensity feature. According to the image edge density, the proportion of the two features can be adaptively calculated to improve the applicability and accuracy of the algorithm.(8) A Mean Shift based random walker interactive image segmentation algorithm is proposed to solve the problems that the objective contour is easily influenced by the natural texture background and computation speed is low. Image is segmented into many small homogeneous regions by Mean Shift pre-segmentation algorithm, and the homogeneous regions are used to build an undirected graph; Color histogram is used as a descriptor to represent the region color feature statistics, and Euclidean distance and Gaussian weighting function are used to describe the similarity of adjacent regions; The discrete potential theory is used to calculate the potential of each node in the graph, and the final image segmentation can be achieved according to the greatest potential of each node in the graph. The results of experiments demonstrate that the segmentation accuracy and efficiency is improved significantly compared with traditional random walker algorithm.(9) An improved random walker algorithm was proposed to solve the problems that the traditional methods for the segmentation and detection of pulmonary nodules can’t segment pulmonary nodules accurately and can’t separation the pulmonary nodules from blood vessels and chest wall. According to the probability that calculated by Dirichlet boundary condition, the image will be divided into three parts:objective region, background region and uncertain region. Euclidean distance was used to calculate the difference betweeny the nodes in the uncertain region and the seed, which can be used to label nodes in the uncertain region. The proposed achieve more accuracy pulmonary nodules segmentation.