Deep Learning Models in Underwater Image Enhancement

Underwater images are of great significance in marine related fields, but the complex underwater environment causes problems such as color distortion and low contrast, which seriously affect information extraction. This study focuses on underwater image enhancement methods based on deep learning, and conducts in-depth analysis and comparative experiments on various models such as FusionEnhance and CycleGAN. Train the model using the EUVP dataset and test it using the UFO-120 dataset, and evaluate its performance using PSNR, SSIM, and UIQMS metrics. Comprehensive evaluation reveals WF-Diff achieves superior performance across PSNR (27.18) and SSIM (0.92) metrics, demonstrating its exceptional capability in structural preservation. Notably, UWCNN maintains advantages in localized denoising (PSNR 24.78), while UGAN excels in holistic quality improvement (UIQMS 3.01).The study also found that different models have their own advantages and disadvantages in quantitative and qualitative analysis, such as WF Diff, which has excellent quantitative indicators but less detail retention than Pix2pix. This study provides important references for the development of underwater image enhancement technology, while pointing out the limitations of current research in model selection, dataset representativeness, and evaluation indicators.