A framework based on the Affine Invariant Regions for improving unsupervised image segmentation

Processing time and segmentation quality are two main factors in evaluation of image segmentation methods. Oversegmentation is one of the most challenging problems that significantly degrade the segmentation quality. This paper presents a framework for decreasing the oversegmentation rate and improving the processing time. Significant variations in both color and texture spaces are the main reasons that lead to oversegmentation. We exploit Affine Invariant Region Detectors to mark regions with high variations in both color and texture spaces. The results are then utilized to reduce the oversegmentation rate of image segmentation algorithms. The performance of the proposed framework is evaluated in decreasing the oversegmentation rate of the well-known Mean Shift method. In conjunction with the proposed framework, we have applied some optimizations on the Mean Shift method to reduce the processing time. In comparison with the original Mean Shift, our experimental results show a twofold speedup and improved segmentation quality.