Research On Contour-based Image Matching Technology

Image matching technology is widely used in the fields of remote sensing, pattern recognition, navigation, medical diagnosis, computer vision and so on. This paper simulated the process of human cognition, and contour-based image matching technology was studied.First, the image was segmented to get complete long contour. Based on the existing method, the determination of approximate dominant color component of color image was improved. Several texture descriptors was analyzed and compared, ultimately the entropy of image was regarded as measurement of texture. Color and texture features of image were integrated to construct four-dimensional feature space, in which the whole pixels were clustered using K-means algorithm. In above process, the number of approximate dominant color component was used to be the number of clustering, and the corresponding color quantization values were regarded as the initial clustering centers. Experiment results show that this method can segment the color image more accurately, adaptively determine the number of clustering. The segment result was consistent with human recognition, and was a good overview of original image, which was beneficial to extract the main contour in the following process.Second, we applied edge detection on the clustering image. Several common edge detection methods were studied, after the comparison, Canny operator was adopted. Then, the contour tracking algorithm was applied to connect the break point and get the sequences of contour points.Finally, contour matching is studied. Contour matching based on curvature and contour matching based on corner points were carried out, after analyzed the shortcomings of these two methods, contour matching algorithm based on Douglas-Peucker was proposed. The contours were approximated using Douglas-Peucker algorithm to overcome the impact of noise and distortion. Then we build a two-dimensional vector for each reserved contour point, which is invariant to similarity transformation and is used as similarity measurement. Experiment results show that this method can achieve correct matching of contour points between four times scaled images, and is robust to rotation, translation. At the same time, this method can also be applied to matching of visible image and infrared image.