Massively Parallel Cellular Matrix Model for Superpixel Adaptive Segmentation Map

We propose the concept of superpixel adaptive segmentation map, to produce a perceptually meaningful representation of rigid pixel image, with higher resolution of more superpixels on interesting regions according to the density distribution of desired attributes. The solution is based on the self-organizing map (SOM) algorithm, for the benefits of SOM’s ability to generate a topological map according to a probability distribution and its potential to be a natural massive parallel algorithm. We also propose the concept of parallel cellular matrix which partitions the Euclidean plane defined by input image into an appropriate number of uniform cell units. Each cell is responsible of a certain part of the data and the cluster center network, and carries out massively parallel spiral searches based on the cellular matrix topology. Experimental results from our GPU implementation show that the proposed algorithm can generate adaptive segmentation map where the distribution of superpixels reflects the gradient distribution or the disparity distribution of input image, with respect to scene topology. When the input size augments, the running time increases in a linear way with a very weak increasing coefficient.