Using Soft Case-Based Reasoning in Model Order Selection for Image Segmentation Ensemble

Abstract

Unsupervised image segmentation is of essential relevance for many computer vision applications and remains difficult task despite of decades of intensive research. In this paper we address two crucial open problems in image segmentation: a problem of choosing a number of regions (k) and a problem of parameter selection, whose solutions depend on the image characteristics and is typically impossible to explicitly define. Thus far, a number of approaches have been proposed to tackle the problems by means of pattern recognition and machine learning. In this work, we propose to incorporate a soft case-based reasoning (CBR) with a multiple image segmentation framework as a new method for automatic selection of k and parameter setting. Soft CBR building upon fuzzy set theory is an efficient approach for handling ill-posed nature of the addressed problems where the underlying models used for solutions are not well understood. A segmentation ensemble framework demonstrates the effectiveness for solving the parameter selection problem by implicitly exploring the parameter subspace and reaching an optimum out of the segmentation ensemble. However, when k is unknown, one way is to impose the objective function to rank on the set of all possible segmentations (which grows exponentially in a number of pixels in an image). Instead of exhaustively searching the space of all possible values of k, we apply a soft CBR to guide the ensemble framework a set of small reasonable k values for each input image, which dramatically increases speed of computation and significantly improves accuracy of a final segmentation result. In fact, our CBR could be incorporated with most general class of segmentation algorithms. Our contribution is a framework for image segmentation that frees the user from the hassles of parameter tuning and model order selection (choosing k).