Image Feature Extraction And Matching Algorithm And Its Application In The Printed Image Detection

This thesis's background is an online system of detecting printed image. The concrete process includes the following steps: printed image's preprocess, image feature extraction and matching. Through the above processing to get the unqualified printed image. This article has mainly done the following work:a) In the feature extraction stage, the Forstner, Harris, SIFT (Scale Invariant Feature Transform) three kinds of feature extraction operators are used on the printed image detection, and contrast with the point's validity, the number of points and the extraction time.Finally, the experiment demonstrates that the SIFT operator is able to satisfy this project's demand much better.b)In the matching stage, for the Forstner and Harris feature extraction operator, firstly, the correlation coefficient is used to accquite the matching points;secondly,the distance constraints are added to reject the false matching points. The experiment done on the same distortion images shows that the Harris operator extract the characteristic points more exactly than Forstner operator in anti-translation, illumination change, but in the dirt nature image and the different content images, both of the algorithms have false matching.During the SIFT algorithm's characteristic match process, the distance constraintsn are added to reject the false matching points after the NN(Nearest Neighbor ). Then,the number of matching points is used to strengthens the match constraints. The experiment demonstrates that the SIFT algorithm extract characteristic point more exactly than Forstner and Harris algorithm in the recognition different content images and the dirt images. Our algorithm can correct matching these images which translated 60 pictures on x-axis and 40 pictures on y-axis,or revolvesd 90 degrees,or had obvious changed on the illumination. Changing the matching threshold value,will satisfy different matching accuracy requirement.A series of image matching experiments demonstrate that had our algorithm had a good matching result on the translation, revolving, illumination change images, as well as the dirt images and the different content images. The experimental results alse prove that our algorithm is feasibility.