Experimental determination of visual color and texture statistics for image segmentation

We consider the problem of segmenting images of natural scenes based on color and texture. A recently proposed algorithm combines knowledge of human perception with an understanding of signal characteristics in order to segment natural scenes into perceptually/semantically uniform regions. We conduct subjective tests to determine key parameters of this algorithm, which include thresholds for texture classification and feature similarity, as well as the window size for texture estimation. The goal of the tests is to relate human perception of isolated (context-free) texture patches to image statistics obtained by the segmentation procedure. The texture patches correspond to homogeneous texture and color distributions and were carefully selected to cover the entire parameter space. The parameter estimation is based on fitting statistical models to the texture data. Experimental results demonstrate that this perceptual tuning of the algorithm leads to significant improvements in segmentation performance.