Enhanced Normalized Cuts with Spectral Weight Adjustment for Image Segmentation

This work investigates an enhancement to the normalized cuts algorithm, introducing a preliminary spectral segmentation analysis to make the process of binary image segmentation more effective. In particular, we propose two improvements: (i) use the results of the proposed preliminary spectral clustering algorithm as a prior for the final segmentation (ii) use Bayesian optimization coupled with Gaussian processes to tunes the hyperparameters. Then, using the results from this procedure, and by maximizing a confidence measure it is possible to obtain a set of pixels belonging to the foreground and the background. These pixels are passed to a min-cut/max-flow algorithm as the source and sink nodes for further refinement, therefore automating the process of foreground/background pixel selection for this algorithm. Finally, the normalized cuts algorithm has been customized to perform video segmentation in real time. Experiments on a subset of the MSRA10K dataset and in different video conditions showed relevant improvements in segmentation quality and real-time usage. The proposed methodologies have the potential to be the starting point for more advanced techniques in fields requiring precise and efficient video segmentation, such as hand tracking.