<p>Underwater image acquisition often suffers from severe color distortion due to light dispersion and absorption. To address this problem, we propose AquaTri-UNet++, a lightweight enhancement network based on an improved UNet++ architecture. We designed a lightweight depthwise residual convolution module (LDR-Conv) to replace the original convolution modules in down-sampling and up-sampling operations. This approach significantly reduces model parameters and computational complexity while preserving the original feature processing capabilities. In addition, we introduced pre-compression in skip connections to further address the substantial memory and processing latency issues inherent in deep convolutional neural networks. Furthermore, AquaTri-UNet++ incorporates a three-stage attention mechanism (Aquatic Triple Attention, AquaTriAtt) tailored for underwater scenarios to learn differentiated transfer parameters, enabling more flexible and robust enhancement. Experimental results demonstrate that the lightweight AquaTri-UNet++ proposed in this paper achieves significant reductions in model parameters, computational complexity, and latency compared to baseline models while substantially increasing frames per second (FPS). Simultaneously, compared to the baseline model UNet++, the enhanced images exhibit improved Structural Similarity Index (SSIM) and peak signal-to-noise ratio (PSNR) by 0.1634 and 6.5491, respectively, demonstrating exceptional real-time processing capabilities. This fully validates the model’s engineering applicability and deployment feasibility in complex underwater optical environments. the enhanced images exhibit improved Structural Similarity Index (SSIM) and peak signal-to-noise ratio (PSNR) by 0.1634 and 6.5491, respectively, demonstrating exceptional real-time processing capabilities. This fully validates the model’s engineering applicability and deployment feasibility in complex underwater optical environments.</p>