Pixel Based Image Segmentation Method Research Using Fuzzy Theory

After three decades of development, digital image processing has becomea very important topic in computer engineering. Image segmentation, often asan initial part of image processing, which usually is a preprocessing step inmany image, video and computer vision applications, always plays a key rolein image processing. It is the most fundamental problem, but also complexand difficult due to the reason that no one common method is suitable for theneeds of different requests.In many practical cases, due to these reasons such as limited spatialresolution, poor contrast, noise, intensity inhomogeneity, digital imaging oftenhas uncertainty and ambiguity characteristics. Fuzzy theory, which is ahotspot in current image segmentation field, is good to describe thesecharacteristics.Based on fuzzy theory, this thesis paper mainly studies fuzzy thresholdingand fuzzy c-means(FCM) algorithm which are commonly used in imagesegmentation, proposes improved algorithm framework and tests in medicalimage processing. The main work and contributions are as following:(1) Propose a fuzzy segmentation framework based on image histogramand pixel similarity. Without a prior knowledge and rather than solving anoptimal problem, this method directly uses fuzzy membership function andsimilarity degree between object area and background area to calculate asuitable threshold, which can avoid getting trap into local minimal. It alsomeans the framework can get a more appropriate threshold selection bychoosing different fuzzy membership function for a series of images.(2) Propose an improved FCM algorithm(mbFCM) based on spatialinformation. Although FCM algorithm and some of its variants have beenextensively used in unsupervised image segmentation applications, they sufferfrom either noise sensitivity or loss of details more or less. This paper presents a novel FCM variation suitable for image segmentation. Theproposed method forms a framework which first makes use of spatialinformation by combining bilateral filtering and furthermore by incorporatingwith multi-resolution, thus providing the following advantages: It is lesssensitive to both high-and low-frequency noise and removes spurious blobsand noisy spot; It yields more homogeneous clustering regions; It preservesdetail, thus significantly improving clustering performance.(3) By the use of synthetic and real image data, this paper also discussesthe mbFCM algorithm used in segmentation of multiple-feature magneticresonance (MR) image. Experimental results and quantitative analysesthrough simulations by MATLAB software based on both evaluationfunctions and ground truth labeling suggest that, compared to other fuzzyclustering algorithms, the proposed method further enhances the robustness tonoisy images and capacity of detail preservation, thus validating theprocessing capabilities in medical image segmentation.