The Research On Image Segmentation Of Natural Images

Ocular organ obtains more than 70% of information among human perception, and all the information is in the form of images, which manifests images play an important role in human perception. As one of the classical problems in computer vision, image segmentation has attracted high attention since 1970 s and being widely used in agriculture, industry, medicine, transportation, meteorology, military, and so on. As the main research object in computer vision, natural color images are commonly seen and used in life, and researches on which are of big value.In this context, the thesis investigates on natural color image segmentation by fuzzy theory. The main contributions are below:Firstly, this thesis analyzed the transformation procedure from RGB to XYZ and to Lab, and deduces that the parameters 0X, 0Y and 0Z in the transformation functions between XYZ and Lab are 100 times of the sums of three lines in the transformation matrix between RGB and XYZ, respectively.Secondly, to solve the noise-sensitive problem of the direct application of PCA on image segmentation, this thesis make use of the spatial information and proposes a new method named PCA-S, which exploits a filtering technique in clustering domain, and thus can effectively recognize and modify the noised pixels, while keeping the edge pixels from being impacted. However, PCA-S is still sensitive to the initial clustering centers, and thus it better be combined with other method which can determine the initial clustering centers.Thirdly, to make MC-FCM more time-efficient and more immune to noise, the thesis takes full advantage of PCA-S and MC-FCM to mutually make up for their shortcomings, and finally proposes a new method named MC-PCA-S. This method uses different thresholds in different subtractive clustering algorithms and introduces memory to the association algorithm, and thus makes it more time-efficient. What’s more, it is proved that this new method is more immune to noise than MC-FCM by simulation.Finally, experiments are done on the natural image database BSDS500, and comparisons between MC-PCA-S and MC-FCM are presented. The mean F-Measure of MC-PCA-S is 0.7007, which is much better than MC-FCM, whose mean F-Measure is 0.5949. This manifests that MC-PCA-S is an effective image segmentation method.