Research Of Image Segmentation Using Phase Information

Image segmentation is the key step from image processing to image analysis. Image segmentation is significant because all subsequent tasks such as feature extraction, object recognition are dependent on the quality of image segmentation.Because of the diversity of objective and the uncertainty of objective imaging in natural image, it is difficult to achieve good segmentation results applying traditional and single segmentation method. Phase Congruency Model avoids the hypothesis of the edge type in advance, and is invariant by contrast. Based on these excellent characteristics of Phase Congruency Model, we use it to detect the edge of natural image. But the original Phase Congruency Model does not consider the direction information of phase data, then "aftershocks" behavior will exist in detected results. To overcome this limitation, Moment Analysis Theory is used to the phase data integration in each direction.However, due to noise interference and the small differences between background and objectives in gray (or color, texture, etc), it is difficult to obtain a closed contour by edge detection. To solve this problem, a new method for image segmentation is proposed in this paper. The method combines the edge detection based on Phase Congruency Model and region growth, and the specific segmentation process is as follows. Firstly, the peak in frequency spectrum of input image is detected using Converging Squares Algorithm (CSA) and the middle scale wavelength of the Log Gabor filters is set by the peak parameter. Secondly, image edges are detected by Phase Congruency Model and the major geometric structures are obtained by removing short edges. Then, after taking the boundary point as the seed pixels and classifying the seeds by K-means, region growth will be carried out in the gradation space or the color space. Finally, small holes are repaired by mathematical morphology method. Experimental results show that the segmentation method proposed in this paper is robust to region texture details, and segmentation results are consistent with the characteristics of human vision system (HVS).