Image Segmentation Based On Semantic Information Of Primal Sketch And Superpixel Merge

Image segmentation is a fundamental and challenging research problem in imageprocessing field. Research on segmentation method should combine both thecharacteristics of the image data itself and the subsequent application of dividing result.In images containing target which has the feature that the body markings have twodifferent colors alternately repeated, like zebras and tigers, to split up the target as awhole is difficult.Based on the visual computing theory, this paper gets the sketch map using theinitial sketch model. Sketch segments characterize the position and direction of thesingularity in the image. In images we are dealing with, singularity information fallsinto two categories, namely, traditional boundaries and the borders between bands andstripes. In image segmentation filed, the traditional boundaries should be reserved asthe final segment boundaries, while the boundary between the bands and stripes is dueto difference in color of adjacent bands. Because of the regularity in bands and stripesof zebra and tiger, in the final segmentation result, different with traditional boundaries,boundary between the bands cannot be the final segmentation boundaries, so as tosegment the target as a whole. Therefore, this paper builds geometric blocks with thesegments that compose the sketch map and then maps the geometric blocks to thecorresponding positions of the original image and extracts the co-occurrence matrixbased on the geometric blocks. We treat the co-occurrence matrix as the features of thecorresponding lines and then divide the sketch segments into band and stripes categoryand general boundary markings category based on the features. For the superpixels gotfrom the over-segmentation method, we merge them under the guidance of thesketch-classification-based semantic information. For the superpixels which be guidedby the band and stripes category, we count up the gray values of them and then make afurther segmentation based on the grayscale statistics symbiotic relationship betweeneach superpixel and its neighbors. Thus, we get the final segmentation result.Simulation results show that the proposed method can get better segmentation results.This paper also applies the proposed method to the color image segmentation.Compared with the gray-scale image, the color image contains data of three channelsand therefore contains more information. Certainly, we can convert the color image togray scale image and then process it. But this will undoubtedly lose much available information. Therefore, this paper gives the three-channel integration policy based onthe analysis on the three-channel of the color images. We fuse the information of thethree channels and then with the proposed method we finally segment the color imagescontaining targets like zebras and tigers. Simulation results demonstrate theeffectiveness of the method.