In this paper, the authors present ConGrap, a novel contour detector for finding closed contours with semantic connections. Based on gradient-based edge detection, a Gradient Map is generated to store the orientation ...In this paper, the authors present ConGrap, a novel contour detector for finding closed contours with semantic connections. Based on gradient-based edge detection, a Gradient Map is generated to store the orientation of every edge pixel. Using the edge image and the generated Gradient Map, ConGrap separates the image into semantic parts and objects. Each edge pixel is mapped to a contour by a three-stage hierarchical analysis of neighbored pixels and ensures the closing of contours. A final post-process of ConGrap extracts the contour borderlines and merges them, if they semantically relate to each other. In contrast to common edge and contour detections, ConGrap not only produces an edge image, but also provides additional information (e.g., the borderline pixel coordinates the bounding box, etc.) for every contour. Additionally, the resulting contour image provides closed contours without discontinuities and merged regions with semantic connections. Consequently, the ConGrap contour image can be seen as an enhanced edge image as well as a kind of segmentation and object recognition.展开更多
文摘In this paper, the authors present ConGrap, a novel contour detector for finding closed contours with semantic connections. Based on gradient-based edge detection, a Gradient Map is generated to store the orientation of every edge pixel. Using the edge image and the generated Gradient Map, ConGrap separates the image into semantic parts and objects. Each edge pixel is mapped to a contour by a three-stage hierarchical analysis of neighbored pixels and ensures the closing of contours. A final post-process of ConGrap extracts the contour borderlines and merges them, if they semantically relate to each other. In contrast to common edge and contour detections, ConGrap not only produces an edge image, but also provides additional information (e.g., the borderline pixel coordinates the bounding box, etc.) for every contour. Additionally, the resulting contour image provides closed contours without discontinuities and merged regions with semantic connections. Consequently, the ConGrap contour image can be seen as an enhanced edge image as well as a kind of segmentation and object recognition.