Perceptual tuning of low-level color and texture features for image segmentation

We perform subjective tests to determine the key parameters of low-level texture and color features for a previously proposed image segmentation algorithm. The parameters include thresholds for texture classification and feature similarity, as well as the window size for texture estimation. The subjective tests use small isolated patches of textures that correspond to homogeneous texture and color distributions. The goal is to determine what information such small image patches convey to human observers, and to relate those to image statistics. We show that this perceptual tuning of the segmentation algorithm leads to significant performance improvements.