A nonparametric statistical analysis of texture segmentation performance using a foveated image preprocessing similar to the human retina

The human visual system is “foveated” in nature. The foveated nature of the human retina can be considered as a visual resource allocation such that there is drastically varying resolution within the field of view. Here, texture segmentation using a retina-like foveated image preprocessor is performed and the texture segmentation performance is analyzed. Latest neurophysiological data from human and macaque retinae are used to determine the parameters of the foveated system. Texture patterns are artificially generated and are passed through the foveated preprocessor. Texture segmentation is done at the output of the foveated module using texture patch classification. At the classification stage, a near-optimal classifier is used with no data reduction or feature extraction stages, so that any change in the segmentation performance is solely due to the effects of visual eccentricity. Texture segmentation performance results with varying eccentricity are obtained. The performance is above human performance and smoothly drops with increasing eccentricity. This study demonstrates the importance of finding the minimum visual resolution which is required to do texture segmentation when a desired performance level is given. `Sufficient resolution classification´ may improve segmentation speeds considerably. From the authors´ experiments, it has also been found out that a strong correlation exists between the human visual performance and the performance of the authors´ artificial foveated visual system