A Panoramic Image Multi-scale Segmentation Method

Machine vision is an interdisciplinary field. With the continuous development of computer science and technology, machine vision become an effective tool for the scholars study visual perception in many field, such as forensic science, medicine, military, remote sensing, weather etc. The rapid development of new sensors improve the ability of acquiring images, so the image information obtained by the single sensor often cannot meet the practical requirements. Human visual system can easily identify the target object, the automatic image segmentation technology is a highly complex integrated problem, especially in the image of the region extraction, industrial cameras can’t perceive the overall information of the object, can only see the local information.To meet this requirement, first of all, to collect the whole image information, and then to get a specific area of the information of target image segmentation. The integrated use of image pretreatment, stitching and segmentation, the researchers can get a clearer, more comprehensive and useful picture detail, Solving image fuzzy, excessive sharpening, noise, information missing and the problems caused by excessive segmentation. Therefore this paper proposes a panoramic image multi-scale segmentation method, The method first using wavelet transform preprocessing, and according the SIFT feature match to stitch panoramic image, and then re-use wavelet transform multi-scale edge detection; To detect the edge information, the combination intensity of multi-scale information can obtain similar the matrix W; The last graph partitioning technology and morphological operation use to get the target. The experimental results show that this method can produce high quality segmentation results, effectively reduces the noise influence, and improves the operation speed.The main research content of this article is as follows:(1) From the spatial, frequency and wavelet domain, the paper discussed the main methods of image preprocessing. Firstly, reviewed the spatial domain and frequency domain for all kinds of classic methods and basic principle of image preprocessing. Then carries on the induction and summary, points out the advantages and disadvantages of various methods and suitable scope. Finally this paper introduces the application of wavelet transform in image preprocessing. It pointed out that the multiresolution analysis of wavelet has good spatial domain and frequency domain localization features, therefore it make use of the wavelet multi-scale edge detection to improve the accuracy of edge contour and segmentation effect.(2) This paper introduces the use of SIFT feature matching of panoramic image stitching algorithm. To scale, first using robust SIFT algorithm was carried out on the processed image feature point extraction and matching.For coarse matching of false match, application consistency random sampling algorithm (RANSAC) screening, finally gradually into a smooth transition method to eliminate the overlap region, complete the image stitching.(3) From a single scale and multi-scale discuss the various image segmentation method respectively. Including point, line and edge detection, threshold processing, segmentation based on region and watershed transform a single scale method and multi-scale normalized cut (Multiscale Ncut) that multiple scale division. According to predefined constraint matrix, multi-scale normalized cut method can under different scale space in parallel, without iteration, can get rough layer and fine layer segmentation of image detail. This method introduced multi-scale space that can greatly make up what due to the weight matrix of graph cut algorithm structure order is too large, and the problem of low computational efficiency.(4) In order to effectively avoid the lost or redundancy of the edge information, and obtain a complete target segmentation, this paper puts forward contour extraction using the wavelet multi-scale edge detection that on the basis of panoramic image. According to the constraint matrix C, edge of the image and the intensity, the weight matrix W can be obtained. Then we can get feature vector from the technology of spectrum segmentation. Finally the discretization and morphology on the open operation, so as to realize the goal of image segmentation. According to the accuracy, recall rate, average absolute error and all aspects of the evaluation standard reveal this method is closer to the standard of artificial segmentation (Ground Truth).