Research And Application Of Superpixel Method Based On Fuzzy Theory

Since the sixties of the last century, image processing has been developed vigorously; the digital image processing technology has been widely used, such as remote sensing satellite images in geography field, CT and MR images in medicine field, images obtained by high-energy particle or an electron microscope in physics field and so on. Image processing has various techniques like image segmentation, image enhancement, image registration, etc. Traditional image processing techniques are in units of pixels and each operation is on the pixel level. The pixel as the basic processing unit contains little useful information, few features can be extracted. Different from the pixel, the concept of a super-pixel is a set as atomic region formed by using some algorithm. The atomic region contains some characteristics that single pixel does not have, for example, shape, boundary contour information and region histogram etc. Such information can be well used in improving the accuracy of the image processing; besides it is well advanced in terms of the time complexity. Thus, in recent years, super-pixel image processing applications become increasingly popular.The traditional super-pixel segmentation methods mostly for natural images, and using two-valued logic to generate super-pixel, When these methods process images with significant ambiguity characteristic, the boundary pixel always cannot be correctly classified, so the processing results are often unsatisfactory. Super-pixel segmented by traditional methods contains a variety of targets, and have low boundary accuracy, this might be not conducive to further image processing.In order to solve the above problems, this paper carried out the work in the following areas:proposed super-pixel method based on fuzzy theory (SMBFT), using fuzzy theory of knowledge as a guide and traditional fuzzy C-means clustering algorithm as a basis, the objective function is been redesigned, using the Lagrange multiplier method to get the approximate optimal solution expression, then through continuous iterative optimization, we can obtain super-pixel segmentation results. This method can make full use of the advantage of the fuzzy clustering algorithm in dealing with processing the images have the ambiguity of the characteristics, make up the disadvantages of traditional segmentation methods in dealing with these images. Boundary pixels which have higher uncertainty can be correctly classified with maximum probability, and have higher accuracy in boundaries. The internal of super-pixel has single and homogeneous goal. Meanwhile, the paper also considers the use of the surrounding neighborhood pixel information to enhance the spatial constraint information, effectively overcome the effects of noise, so that the proposed algorithm has better robustness.The algorithm selected the Berkeley database images and brain MR images from the Brain web to do the verification experiment. And this paper proposes a comprehensive criterion to measure the weight of the two kinds of criterion in choosing super-pixel methods for color image. Evaluation criteria for medical image data sets is using internal entropy of super-pixel which inspired by the concept of entropy in information theory. The experimental results show that this method has more obvious advantages both in natural images and medical images than traditional methods, demonstrate the effectiveness and universality of this method.