Refining Image Segmentation by Integration of Edge and Region Data

An Iterative Parallel Region Growing (IPRG) algorithm has been previously developed.% produces hirer$rchich scpntations of images from finer to coarser resolution. Then, depending on the application, the "best" segmentation may be chosen as given by one particular iteration. But an ideal segmentation does not always correspond to one single iteration but to several different ones, each one producing the "best" result for a separate part of the image. With the goal of finding this ideal segmentation, we chose to refiie the results of the IPRG algorithm by utilizing some additional information, such as edge features, and by interpreting the tree of hierarchical regions. REGION SEGMENTATION Image segmentation is the process by which the individual image pixels are grouped into partitions, according to some intrinsic properties of the image, e.g. grey levels, contrast, spectral values or textural properties. The segmentation results are then used at a higher level to interpret the image. Segmentation algorithms are usually described as edge-based or region-based segmentations. One or the other of these techniques are utilized depending on the data and the application, but both types of measurements may be altered by different factors; for example edge detection may be very sensitive to noise while texture measurements may change under non-uniform lighting. Thus by combining region extraction and boundary detection, the results of both independent segmentations may be improved, as one technique may correct the "errors" of the other. For example, broken contours are common to many edge detectors, while region segmentation always produces closed contours. On the other hand, edges may help in choosing the regions of interest among all the regions provided by a region-based algorithm. This fusion can be done either at the pixel or at the symbol level. The drawbacks of data fusion at the pixel level are the large amount of data to process and the registration of disparate information; but at this level, all meaningful information is still present. On the other hand, at the symbol level, the amount of data to process is reduced, the registration problem is easier to resolve, but some meaningful information may have been lost in the extraction of the symbols. If the two processes are integrated in an iterative fashion such as a region growing approach [Ti189], the distinction between the two levels of fusion disappear since this approach starts at the pixel level to iteratively grow the regions; and the computational problem is resolved by utilizing a parallel algorithm. UTILIZING EDGES FEATURES TO REFINE A REGION SEGMENTATION The Iterative Parallel Region Growing (IPRG) algorithm has been previously developed and implemented on a SIMD ("Single Instruction Multiple Data") architecture, the MasPar [Ti189]. It produces hierarchical segmentations of images from finer to coarser resolution, by utilizing some similarity criteria that are function of the original image data; e.g., the normalized mean square error of the grey level variance or the change in image entropy. At each iteration, a set of subimages is defined Then, for each subimage, the most similar pair of spatially adjacent regions is merged. Figure 1 shows on a test example how the regions grow from one iteration to the next. Starting from the same original image, the edges are computed by a Canny edge detector [Can861 and the parameters are adjusted in such a way that the final edge image contains only "good" edges, although these edges do not necessarily form closed contours. The parameters can be chosen either automatically within the program or interactively by a user. By "good" edges, wemean theminimum amountof edges necessary for thefinal James C. Tilton Mail Code 930.4 NASA Goddard Space Flight Center Greenbelt, MD 20771 Tel: (301) 286-9510 Fax: (301) 286-3221 interpretation. Figure 2 shows some "good" edges corresponding to the test example of Figure 1. lteratlon A lteratlon B lteratlon C p$q