Traditional image segmentation methods based on superpixel still have many problems in terms of consistency of edge segmentation, computational efficiency and adaptability of merging algorithms. We combine domestic and foreign research advances and propose a novel superpixel merging image segmentation method, which adopts ERS superpixel over-segmentation algorithm and uses intensity and gradient histogram as superpixel features. Additionally, EMD method is used to calculate feature distance and the merging self-adaptive threshold is obtained by mixing Weibull model to complete the segmentation. As a result, the time complexity of proposed algorithm is reduced to O(N), and the segmentation process is not required to manually select the region to be segmented. Compared with current methods, experiment results show that the proposed method has better performance on boundary accuracy and processing efficiency.