| Compared with the text,the image is more explicit,visual and situational,it is one of the important ways for human to obtain information.The development of science and the changes of users' habits leading to that there are massive image information on the Internet.Therefore,when studying the targets of users' interests,we need to process the massive image information and pick up effective ones.The image processing could be divided into several process such that image denoising,image segmentation,image recognition,image fusion.The image segmentation is an important part of the image processing,how well of the segmentation results affects the subsequent processes.As a preprocessing step of image processing,the image segmentation is an important research topic.The segmentation method is mainly based on pixel and super-pixel.This thesis mainly studies the segmentation method based on super-pixel.The super-pixel of the image is the region composed of the consistent adjacent pixels such as color,texture,shape and so on.Using super-pixel instead of pixel to represent image can reduce the basic element numbers and redundancy of image greatly,which reduces the operation scale of processing such as subsequent pattern recognition,target tracking,image searching.The main contents of this thesis are as follows:(1)The traditional k-means algorithm is improved,and a super-pixel generation method based on boundary matching,which named BM-k-means is proposed.Due to the complexity of the natural information contained in the image,the super-pixels generated by the existing algorithms are not ideal enough,and their features will restrict and affect each other.Based on the traditional k-means,this thesis use Mean Shift to initial seed points,use 3δ principle to filtrate super-pixels of high credit worthiness,in order to update the seed point.Introducing boundary items,data item and firming concept among the similar distance function,proposing a super-pixel generation method based on boundary matching,which named BM-k-means.The super-pixels generated by BM-k-means is more ideal than the traditional k-means.(2)Improved the traditional DBSCAN algorithm and proposed a new clustering algorithm,named Grid-DBSCAN.Because the traditional DBSCAN will search the full-graph when clustering each point,the search size is proportional to the image size.Each point will only be marked once that could easily lead to segmentation errors caused by the failing of updating the class label.In this thesis,the concept of grid division and minimum accessible distance is introduced in the traditional DBSCAN algorithm,and a new clustering algorithm is proposed,which reduces the search size and makes the class labels of each point modifiable.(3)Combining the above BM-k-means and Grid-DBSCAN algorithm to generate super-pixels and clustering the super-pixels,the final segmentation results are obtained.The image segmentation algorithm proposed in this thesis is compared with the traditional SLIC,Mean Shift and Graph-Based algorithm on the BSD500 data set,and its rationality is verified.Experiments show that the segmentation effect of the proposed image segmentation algorithm is better than the traditional algorithms. |
