Abstract: To address the issues of color deviation, contrast reduction, and detail blurring in underwater images caused by light refraction and absorption, this paper proposes an underwater image enhancement method with color space optimization based on a lightweight U-shaped network（DU2 Net）. Firstly, the U-shaped network is optimized leveraging a large-scale dataset of real underwater scenes（DSUI） containing 11 739 images, along with high-quality reference images, semantic segmentation maps, and medium transmission maps. Axial depth-wise convolution and dense attention blocks are employed to effectively reduce computational complexity and parameter quantity, thereby significantly enhancing DU2 Net processing speed and image enhancement quality. Secondly, a multi-color-space loss function combining RGB, LAB, and LCH color spaces is introduced to better align with human visual perception characteristics, further improving the color fidelity and contrast of the restored images. Experimental validation demonstrates that compared to state-of-the-art underwater image enhancement techniques such as UDCP and CRUHL, DU2 Net achieves improvements of 0.367, 0.072, 26.165, and 7.833 on the UIQM（underwater image quality measure）, UCIQE（underwater color image quality evaluation）, CCF（color cast factor）,and AG（average gradient） metrics, respectively. Furthermore, DU2 Net attains a processing speed 8 times faster than UDCP.These results validate the applicability and efficacy of the proposed method across diverse underwater scenarios.