| Summary: | 碩士 === 大同大學 === 資訊經營學系(所) === 98 === This study proposes an effective method for image retrieval in case that images are magnified or reduced. In contrast to the conventional keyword-based approach, the method makes use of image shapes. We determine interested points in the curvature-scale space (CSS) to recognize similar image objects. Images are processed in which edges are first extracted from original images and, then, corners are determined by the change of curvature along the edge. The corners are connected as vectors. Relative positions of the corners are determined by transposition and normalization. Based on these corners, we select the corners with similar curvature and calculate their relative distance. Similarity of two images is then the summation of the proportion of all relative distances. Similarities among images are ranked to determine which image is most relevant for a given image. To evaluate the effectiveness of the method, we use an image magnified by a factor from 0.2 to 3.0. The proposed method incorporated with a web crawler is expected to realize an automated image recognizer.
|
|---|
