Research On Image Segmentation Methods Based On Superpixel

Image segmentation is one of the most important research fields in image processing, because segmentation results have a big impact on the subsequent tasks, such as image analysis and image understanding, etc. Superpixel segmentation methods are image pre-processing technologies that have rapidly developed in recent years. These methods can segment an image into a certain amount of partially semantically meaningful sub-regions in a short time. Compared to the basic elements-pixels in traditional image processing methods, superpixels have the advantage of better abstraction of image local features. Furthermore, superpixels can dramatically reduce the complexity of the subsequent processing. Given these significant advantages, superpixels have been widely used in many computer vision fields, and have become a hot research topic in image processing in recent years.At present, the commonly used superpixel segmentation algorithms have a common limitation of poor segmentation qualities with a small number of superpixels. However, given the complexity of the subsequent processing while superpixels used as segmentation primitives in traditional algorithms like region merging, it is preferable to keep the number of superpixels to a minimum. It is of practical value to improve the superpixel segmentation quality with a small number of superpixels. In the application of quality test of Printed Circuit Board(PCB), the accurate segmentation of via holes from the PCB CT(Computed Tomography)image is needed for the subsequent processing tasks like solder joint quality test and circuit analysis. However, the accurate segmentation of via holes is difficult because of the characteristics of much noises and the low contrast between the object and background in PCB CT image. It is helpful to segment via holes with superpixels as segmentation primitives, due to the superpixels’ abilities of anti-noise and representations of image local structural information.Aiming at the limitation of superpixel segmentation algorithms and the via holes segmentation problem of PCB CT image, this paper studies superpixel segmentation algorithms and the applications of superpixels in image segmentation. The main works are presented as follows:The basic ideas and segmentation performances of commonly used superpixel segmentation algorithms are analyzed, and the mainstream methods for superpixel quality evaluation, compactness evaluation and the algorithm controllability evaluation are presented. Then the qualitative comparison results and the summary of the algorithm controllability of sixteen segmentation algorithms are given. Six typical algorithms are compared in view of segmentation performances according to the experimental results of quantitative evaluations. The interrelations between the superpixel qualities and the number of superpixels are summarized as well as the interrelations between the superpixel compactness and the number of superpixels. Finally, we sum up the existing limitations of superpixel segmentation algorithms.This paper proposes a hierarchical superpixel segmentation model. Aiming at the common limitation of poor segmentation qualities with a small number of superpixels, a hierarchical segmentation structure containing the coarse segmentation stage and the fine segmentation stage is applied in the proposed model, which can adjust the superpixels distribution density according to the local structural characteristics. Experimental results show that the proposed model greatly improves the segmentation quality in a small number of superpixels compared with traditional superpixel segmentation algorithms. Therefore, a region merging image segmentation algorithm based on the hierarchical superpixels is proposed. Aiming at shortening the iterative merging time of the MSRM(Maximal Similarity Based Region Merging) algorithm, the stages of superpixel segmentation, user interaction and feature extraction and similarity measurement are all improved in our method. Compared with the MSRM algorithm, our method keeps the same segmentation quality with a significantly shortened merging time, improving the segmentation quality compared to other algorithms like Grabcut, etc.A superpixel-based segmentation algorithm for the via holes in PCB CT image is proposed. First, the via holes segmentation results of seven typical segmentation algorithms are shown. In order to choose the proper superpixel segmentation algorithm for PCB CT image, our method studies the segmentation performances of ten superpixel segmentation algorithms, in which the ERS(Entropy Rate Superpixel)algorithm is the best. A region-merging-based superpixel merging and pruning strategy is designed to extract the via holes from the backgrounds. Experimental results show that compared to the state of the art segmentation algorithm based on the circle detection of Hough transform, our method has a certain degree of improvement not only in the segmentation Precision Rate metric but also the Recall Rate metric, especially in reducing the error probability when facing the low contrast or gray-level overlapping between the via holes and the backgrounds.