Bağlanırlığa zorlanmış bayesçi süperpikseller.

In this study a general flow for clustering-based superpixel (SP) extraction methods is presented, while each step is analyzed in detail, and improvements are proposed. Considering general SP extraction method steps, initial grid alternatives are examined. The necessity of initial grid refinement is studied and unlike current approaches, a novel Edge Based Refinement step which does not break regular grid structure is proposed. Label update constraints are also analyzed in terms of preserving regular initial tiling, and Just Connected method enforcing connectivity from the beginning is proposed. The requirement of adjusting one of hyper-parameters, iteration count, for different image resolutions and different number of SPs is eliminated by determining the number of iterations relative to SP area. Considering these proposals,extensions to the state-of-the-art SP methods, SLIC+ and LASP+, are proposed which normalize spatial term with SP spatial covariance. Novel cost-functions SLIC++ and LASP++ are also presented for a further improvement with the normalization of spectral term with SP specific dynamic parameter. Finally, a Bayesian classifier is proposed for pixels during SP label assignment. Based on improvements in various steps mentioned above, a family of superpixel extraction methods including, SLIC++/R, SLIC++/H, LASP++/R. LASP++/H, BSP/R and BSP/H are presented. For the evaluation, a novel Boundary Achievable Segmentation Accuracy metric is proposed that replaces three frequent metrics from the literature. Compactness and area under curve approaches are also proposed as evaluation methods to minimize any performance ambiguity for the literature benchmarks. Both proposed spectral term and spatial term improvements significantly increase accuracy of generated SPs with no execution-time burden. In addition, employing Bayesian classifier leads to generate more accurate SPs in a shorter amount of run-time. Besides, with the proposed label update criteria, connectedness of SPs are ensured during generation process that preserves regular grid topology enabling them to be fed into conventional neural-networks.Thesis (M.S.) -- Graduate School of Natural and Applied Sciences. Electrical and Electronics Engineering