Fast and Effective Image Segmentation via Superpixels and Adaptive Thresholding

This paper proposes a fast and effective image segmentation algorithm by firstly clustering image pixels into a small number of superpixels and then merging these superpixels whose distances are below an adaptive threshold together to get the final segmented fields. The adoption of superpixels dramatically decreases the computation cost, while the adaptive thresholding aims to select a reasonable segmentation from a set of possible segmentations with hierarchical scales. The adaptive threshold can be calculated with a fast sequential procedure. Experiments on Berkeley Segmentation Data Set (BSDS500) demonstrate that our proposed algorithm is competitive to other state-of-the-art segmentation methods. Moreover, this segmentation framework can be improved to excellent performance by using more elaborate superpixel algorithms.