Image Segmentation Through Automatic Fractal Dimension Classi ̄cation

Usual segmentation techniques of grayscale images depend on supervised trial-and-error procedures. Moreover, in noisy images, local classi ̄cation schemes fail due to the random °uctuations introduced by the noise. Recent proposals as the active contours may be robust enough to cope with some cases of noisy images without supervision (except for the initialization step), but still fail with images with non additive noise that are common in remote sensing, SAR images and other cases. In this work we propose a di®erent approach to noisy image segmentation, based on fractal dimension classi ̄cation. Instead of detecting local changes in the image grey level, a previous step of local fractal (box counting) dimension is applied. Histogram classi ̄cation of this attribute of the image allows an adequate threshold detection. Therefore, the complete segmentation procedure is fully automatic. The segmentation results obtained with this classi ̄cation procedure are remarkably robust, even with images with non additive noise.