Texture Relative Superpixel Generation with Adaptive Parameters.
Xiao Pan,Yuanfeng Zhou,Zhonggui Chen,Caiming Zhang
DOI: https://doi.org/10.1109/tmm.2019.2895498
IF: 7.3
2019-01-01
IEEE Transactions on Multimedia
Abstract:Superpixel generation, which is an essential step in many image processing applications, has attracted increasing attention from researchers. In this paper, we present an efficient flooding-based superpixel generation algorithm that generates compact and highly boundary adherent superpixels. In particular, by considering various superpixel properties, we measure the similarities between image pixels by proposing a new distance metric that combines various image features (e.g., colors, spatial locations, neighbor information, and texture features). To control the relative significance of these image features, we acquire the weights of image features (e.g., colors, texture features, and neighbor information of pixels) through a neural network. Then, the final superpixels are obtained through a greedy optimization that considers both the current superpixel and its neighboring superpixels. We perform extensive experiments on two datasets to verify the efficacy of our algorithm. The results show that our algorithm has considerable advantages over existing state-of-the-art methods, particularly regarding the compactness of the resulting superpixels.