Semantic image segmentation aims to partition an image into separate regions, which ideally corresponds to different real-world objects. Many segmentation algorithms have been proposed, exploiting a wide variety of image features and characteristics. It has been shown through empirical studies that segmentation methods that assume a good segmentation partitions an image into different homogeneous regions are likely to fail in non-trivial situations, while methods based on perceptual organization generally generate more favorable segmentations. Yet no formal justification has been provided. In this paper, we propose an information measure for images, the perceptual information, based on human visual perception organization. Using perceptual information, we justify that homogeneity-based segmentation methods are inherently biased, and by incorporating knowledge, perceptual organization can overcome the bias and generate better segmentations.
展开▼
