| Summary: | 博士 === 國立臺灣大學 === 生物產業機電工程學研究所 === 94 === In this study, we developed a synergistic approach for the segmentation of touching ellipse-like objects with obvious texture and noises in an image. The proposed approach modifies and integrates several major image processing methods including pre-filtering, grouping, creating initial contours, and reconstructing contours. For de-noising, mean shift algorithm and Gradient Vector Field (GVF) are employed as a pre-filter. Through the filtering and edge detection, the processed image only preserves the boundaries of objects and rejects noise. With the edges, we developed two kinds of active point grouping approaches for generating the field center of each touching ellipse-like object. Inverse GVF (IGVF) field and mean shift algorithm with distance transform (DT) weight map are employed in the two grouping approaches, respectively. For creating initial deformable contour of each object, we designed two generation methods, the equally-spacing active points method inspired by Monte Carlo’s concept as well as Fitzgibbon’s optimal ellipses method. Finally, the complete contour of each object could be correctly reconstructed by Active Contour Model (ACM). The result shows that the algorithm could successfully reconstruct the whole contour as long as more than 50% of piecewise edge information remained in an image. Compared with the original contours, the ones generated in this study achieved more than 96% similarity. When the obvious textures or noises are filtered out by the mean shift algorithm with GVF weight map, it could effectively remain the edges of the detected objects. Even for an image polluted by 10% salt and pepper noises, the approach still can effectively and robustly eliminate the added noises. Therefore we can successfully cluster objects and reconstruct their corresponding contours by applying active contour model approach. The complete contours of touching objects could facilitate the subsequent image processing to obtain the geometric, texture, and color characteristics of objects in an image. These features might then be used for further clustering, classification, or image understanding.
|
|---|
