Pixel value ordering (PVO) is one of the most commonly used methods in reversible data hiding (RDH). Most PVO methods currently chunk images, categorize them according to the space complexity of the chunks, and adopt different embedding strategies for different chunk categories. Although this approach achieves some results, the utilization of pixel redundancy is still insufficient. In order to expand the embedding capacity of RDH and improve the embedding performance, a high-fidelity reversible data hiding method based on pixels selection and recursion-PVO is proposed in this paper. Specially, for uniformly divided non-overlapping blocks, the set of pixels with the smallest and second-smallest variance values within each block is selected to form a new block to be embedded, thus further utilizing pixel redundancy. Then, the embedding capacity is improved by transforming some reference pixels into to-be-embedded pixels and forming pixel pairs with other reference pixels to calculate the prediction error. Experimental results show that the proposed method has better performance compared to some state-of-the-art methods. In particular, for the commonly used test image Lena, the PSNR after embedding 10,000 bits of data is up to 63.13 dB.