Research On Automatic Segmentation Of Color Images Based On Mean Shift And GrowCut

Image segmentation is a technique that divides an image into meaningful sub-regions.For example,an aerial photo can be divided into specific areas such as residential areas,forests and lakes.In recent years,the pursuit of intelligent color image segmentation has become a new hot topic in image segmentation.Because the color image carries more information than the grayscale image,the existing single image segmentation method can not meet the actual requirements of image segmentation.Therefore,it is necessary to propose novel and more effective color image segmentation methods based on the original image segmentation method,combine the advantages of image segmentation methods and apply it to the field of color image segmentation.This thesis introduces related theory,segmentation methods and evaluation criteria in the image segmentation field.The thesis focuses on the adaptive mean shift and GrowCut automatic color image segmentation algorithm.The main work is as follows:(1)For the problem that traditional mean shift algorithm can easily destroy the integrity of target area and equally treat the sample in the process of color image segmentation,This thesis proposes an adaptive bandwidth mean shift color image pre-segmentation algorithm.The traditional mean shift color image segmentation method can not adaptively select the bandwidth according to the features of image pixel color and spatial distribution.This thesis uses the knowledge of near-neighborhood of pixels to automatically calculate the image segmentation adaptive bandwidth based on the research of traditional mean shift image segmentation,and use standardized Euclidean distances to improve the shortcomings of the mean distance algorithm's Euclidean distance.Experimental results show that the proposed algorithm has a better segmentation effect than traditional mean shift algorithm in image segmentation.(2)For the GrowCut algorithm image segmentation requires the user to interact with the seed mark,this thesis proposes a seed point auto-location method that combines adaptive mean shift algorithm and saliency detection.The subjectivity and uncertainty of traditional artificial seed points mark will directly affect the iteration speed and segmentation effect of GrowCut algorithm,and it is difficult to achieve fast and accurate automatic segmentation.In this thesis,the adaptive bandwidth mean shift algorithm is used to pre-segment the color image directly,and the spatial similarity regions of the merged image are clustered to enhance the image characterization.Thenthe seed region is automatically marked by the saliency detection and morphological operation methods to improve the seeds efficiency.Experiments show that this method overcomes the limitations of artificial image seeds mark,makes full use of the pixel space and color information of color images,and improves the efficiency and accuracy of color image segmentation.(3)In this thesis,the proposed method is firstly evaluated by the analysis method from the perspective of algorithm principle and segmentation performance,compared with traditional GrowCut,fixed bandwidth mean shift(FMS),adaptive bandwidth mean shift(AMS)and mean shift normalized cut(MSNCut).Then the segmentation results are compared by experimental methods.Experiments show that the method proposed in this thesis fully considers the integrity of the segmentation of color image target area,and can achieve automatic segmentation of color images.The segmentation result is satisfactory,and the proposed method effectively validates the validity and accuracy of the proposed method.