Acceleration of clustering-based superpixel algorithms with low memory costs

As a pre-processing step of image segmentation, superpixel algorithms are used to produce small, uniform and compact regions, which can be used for region-based image coding, region-based image processing, and object recognition. In order to meet the requirements of real-time applications for embedded computing, it is necessary to reduce the computational costs of superpixel algorithms and increase the processing speed. In this paper, a series of acceleration schemes for superpixels algorithm is proposed. The features and contributions of this work are stated as follows. Firstly, the spatial distances and the color distances are calculated individually, so that the redundant distance computations can be saved. Secondly, by searching the nearest cluster centroids with centroid priority, the nearest clusters can be found at an early stage. Thirdly, the early-termination mechanism can be applied to the search process to speed up the algorithm without decreasing the quality of image segmentation. Fourthly, the storage for label images and distance images is not required since the operations of nearest centroids are processed in the inner loop of the algorithm. The experiments show that the proposed method achieves the same level of performance as the related work with only 75% of distance computations and 33% of memory costs.