Currently, the supervised models trained on synthetic paired datasets have weak generalization ability and perform poorly in diverse real underwater environments. Although unsupervised models are not dependent on paired datasets, the lack of feature information may result in the generated images with poor visual quality. Therefore, with the architecture of cyclic generation adversarial networks, the underwater image enhancement method of multi-feature selection and bidirectional residual fusion is proposed. On one hand, a multi-feature selection module based on mixed attention is designed to select multiple features of underwater images. Furthermore, the bidirectional residual fusion is used to optimize traditional U-shaped skip connection, which realizes high-efficiency expression of image features and effectively restores the texture and color of underwater images. In addition, mixed attention is introduced and content-aware loss and style-aware loss are proposed in the discriminator to ensure that the enhanced image is consistent with the clear image in terms of global content, local texture, and style features. The PSNR of the proposed model is improved by 6% and 2%, respectively, compared with the existing unsupervised and supervised models. Additionally, SSIM is improved by 4% and 3%, respectively. With a significant enhancement effect on underwater images, the proposed method demonstrates superiority over other existing methods in terms of color fidelity and saturation.