Research And Application On Several Issues Of Image Segmentation

Image segmentation is a basic and important problem in image processing, and it is also the foundation of the researches on image analysis and computer vision. Although, some effective algorithms have been proposed in recent years, there are some deficiencies and problems that needed to be improved in most researches. This paper deeply investigates image segmentation according to different segmentation features, and the primary works and remarks are as follows:(1) For the threshold segmentation, this paper presents a new threshold segmentation measure based on the multi-properties in ultra-fuzzy set. The algorithm solves the segmentation problem in ultra-fuzzy set, and obtains the optimal threshold by comprehensive assessment function constructed by ultra-fuzzy entropy and ultra-fuzzy similarity, which evaluate performances of the segmentation operation from different aspects. Comparing the ordinary thresholding algorithms, the proposed method embodies better results.(2) For the edge detection, we propose a new edge detection algorithm based on edge classification and Hopfield neural network optimization. Firstly, a new edge detector is employed to label the edges, which is improved for overcoming the miss-detection of abrupt and end of the edges, isolated noise. And then, the Hopfield neural net is applied for enhancement of the labeled edges by recovering missing edges and eliminating false edges. In experiments, this algorithm is proved that it can obtain the moderately thick edges for the image.(3) This paper also presents two contour extraction algorithms, which are based on visual perception theory and interactive segmentation respectively. The visual perception based model effectively suppresses the interference of the background texture and enhances the response of the neurons with the same configuration, and obtains a clear contour by simulating the inhabitation and enhancement of cells in primary visual cortex. Moreover, the improved random walk algorithm embodies the better structure features of an image by proposing a new adaptive anisotropy structure tensor to replace the intensity to represent the weights in common random walk algorithm. Both of them overcome the interference of local texture information in the segmenting process. The experiments show that two extracting methods can all obtain the desired results.(4) Propose an unsupervised algorithm for automatically extracting the salient object by combining the global feature and local feature of the image based on visual perception theory. The method firstly obtains the global feature by simulating the cell responses of different scales and different orientation. And then, define the local feature according to the unique features of the salient object as supplement. So, the salient object in an image can be located exactly.(5) To locate the vehicle license palate accurately, a new detecting measure is proposed in this paper. The method presents four feature operators by analyzing the unique features of vehicle license palate to characterize vehicle license palate. And, combine them with morphology and self organizing map neural network to achieve the accurate location. The experiments prove that the proposed measure can overcome the deficiencies of the common locating algorithms, and the total correct recognition rate is 96.25% according to 80 test images.