We have designed an image segmentation scheme which is appropriate for large images and database retrieval applications. The proposed scheme utilizes the Gabor texture features extracted from the image tiles and performs a coarse image segmentation based on local texture gradient. The following figures shows the different stages of the segmentation algorithm, which are summarized as follows.
Local texture flow computation: Using the feature vectors, a local texture gradient is computed between each image tile and its surrounding eight neighbors (Figure a). The texture gradients in the same orientation are combined together to form a texture flow. This texture flow contains information about the direction and energy of the local texture boundary. Via the interactions between different orientations of local texture flow, the direction for finding salient image boundaries can be identified. Figure(b) shows the results of the flow interaction where the vector sum of remaining texture flow is represented by the arrows whose directions point to the texture boundaries, with darker intensities representing stronger texture gradient.
Following the orientation competition, the local texture flow is propagated to its neighbors if they have the same directional preference. The flow continues till it encounters an opposite flow in the image boundary. This helps to localize the precise positions of the boundaries and concentrate the edge energies along the image boundaries. Figure(c) shows the results of this stage.
After the propagation reaches a stable state, the final texture flow energy is used for boundary detection. This is done by turning on the edge signals between two neighboring image tiles if their final texture flow point in opposite directions. The texture edge energy is then defined to be the summation of texture flow energies in the two neighboring image tiles. Figure(d) shows the results of this stage.
The previous stage results in many discontinuous image boundaries. They are connected to form an initial set of image regions (Figure(e)). At the end, a conservative region merging algorithm is used to group the similar neighboring regions. The final image segmentation result is shown in Figure(f). Figure(g) shows a segmented aerial photo.
These materials are presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each authors copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder. W. Y. Ma and B. S. Manjunath, "Edge flow: a framework for boundary detection and image segmentation," Proc. IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), San Juan, Puerto Rico, pp. 744-749, June 1997. [abstract]