A Region-based SRG Algorithm For Color Image Segmentation

In this paper, we present an automatic seeded region growing (SRG) based region algorithm for color image segmentation. The method uses regions instead of pixels as the seeds of SRG. The architecture of the algorithm can be described as follows.First, in the pre-processing stage, the input RGB color image is transformed into HSI color space, because HSI color system has a good capability of representation of the colors of human perception and is a commonly used color space in image processing. In the second step, we use watershed algorithm for segmentation of a given color image. Watershed algorithm acts as one of the most powerful tool for image segmentation. However, the result of watershed algorithm is usually an over-segmentation image because of the noise and other factors. Fortunately, the regions of the over-segmentation can be qualified as the initial seeds for SRG. Therefore in the third step we introduce the automatic seeds selection process from the regions obtained from step two. In step four, region-growing procedure is called to acquire the regions. And in step 5 regions merging procedure is called to merge similar or small regions.We can compare our algorithm and traditional SRG algorithm here. First, our method use regions as initial seeds rather than pixels. In this sense, high-level knowledge of the image partitions can be exploited through the choice of the seeds much better because region has more information compared to pixels. Second, our method is more efficient. The most time-consuming step of the segmentation is the process of region growing. For traditional method, the time complexity is O(nlogn), in which n denotes the number of pixels of original image. While for our method, the time complexity is O(mlogm), where m denotes the number of regions produced by watershed algorithm. Obviously m is much smaller than n especially for very large images.We have implemented the algorithms in MATLAB. Experimental results have shown that our algorithm can produce excellent results.