UNDERWATER IMAGE ENHANCEMENT USING A REDUCED‑DEPTH U‑NET:

Abstract

Background: Underwater images are commonly affected by wavelength-dependent light absorption and scattering, which produce colour distortion, reduced contrast and haze-like degradation. The aim of this study was to design, implement and evaluate a supervised deep-learning system for underwater image enhancement based on a reduced-depth U-Net encoder–decoder architecture.
Methods: The network was trained end-to-end on paired underwater imagery from the EUVP (Enhancing Underwater Visual Perception) dataset, obtained through a Kaggle mirror, using an L1 reconstruction loss and the Adam optimizer, without data augmentation, batch normalization, dropout, residual connections or adversarial training. A fixed-seed 80/10/10 split produced 9,149 training, 1,143 validation and 1,143 test pairs. Performance on the held-out test split was measured with two full-reference metrics, Peak Signal-to-Noise Ratio (PSNR) and the Structural Similarity Index (SSIM), and with a custom no-reference metric adapted from the Underwater Image Quality Measure (UIQM).
Results: The system achieved an average PSNR of 29.36 dB, an average SSIM of 0.8610 and an average custom UIQM-style score of 3.5692 on the 1,143-pair test split. The trained model was additionally deployed through an interactive Streamlit application supporting image upload, adjustable-resolution inference, before/after comparison and no-reference quality feedback.
Conclusions: A transparent, purely supervised reduced-depth U-Net pipeline, with a deterministically defined dataset organization, quantitative evaluation and an accessible interactive interface, provides a practical and reproducible basis for underwater image enhancement and for further development.