Adaptive Superpixel Segmentation With Non-Uniform Seed Initialization

Abstract

Superpixel segmentation is a powerful image pre-processing tool in computer vision applications. However, fewer superpixel segmentation methods consider automatically determining the number of initial superpixels. Focusing on high-precision and connectivity, we propose a superpixel segmentation algorithm with non-uniform seed initialization. The proposed algorithm can adaptively determine the number and the position of initial seeds, and is robust to the segmentation of small objects and slender regions. First, we propose a seed initialization scheme based on the side of the circumscribed rectangle of the small object and interval boundary gradient. To enhance the regularity of superpixels, we equally added seeds for grids with sparse seed distribution. Second, we construct a weighted distance measure with search region and feature constraints, which reduces the computational complexity and enhances the precision of pixel label assignment. Finally, we quantify the disconnected regions are present in abundance, and propose a post-processing method based on the area of predefined small objects. The proposed method can significantly improve the connectivity and regularity of the generated superpixels. Extensive experiments on the widely-used BSDS and CamVid datasets demonstrate that the non-uniform seed initialization is effective, and the performance of the proposed superpixel segmentation is favorably compared with the state-of-the-art methods.