This paper propose a new unsupervised image texture segmentation algorithm based on the design of general optimal FIR filters.
A pyramidal decomposition divide the image into equal dyadic regions/cells. A new similarity function is developed to compare adjacent cells. Similar cells are equally labeled through a labeling process yielding a number of distinct areas. Optimal filters are then computed with respect to each area. The final segmentation result is obtained with a local energy function followed by the c-means algorithm. Recommandations for further improvement of the algorithm are supplied.
Experiments shows that general optimal FIR filters are capable to produce satisfactory segmentation results in an unsupervised texture segmentation system.