Adaptative evaluation of image segmentation results

We present in this article a new unsupervised evaluation criterion that enables the quantification of the quality of an image segmentation result according to the type of the original image. We first briefly present a comparative study of existing unsupervised evaluation criteria. Then, we present a method for the determination of the type of the original image: uniform, mixed or textured by using a learning method (support vector machine). In the third part, we present the proposed algorithm for segmentation evaluation and the experimental results on synthetic images from a large database. Last, we conclude and present some perspectives of this work