Underwater Image Enhancement Using Deep Learning: A Multi-Stage Processing Approach

Abstract

Capturing images beneath the water surface is fundamentally different from photography in air. Water selectively absorbs different wavelengths of light, scatters photons through suspended particles, and strips images of natural colour, contrast, and sharpness before they ever reach a sensor. The result is a degraded visual record that makes downstream tasks — coral reef surveys, pipeline inspections, autonomous vehicle navigation, and archaeological documentation — significantly harder than they need to be. This paper presents a systematic, multi-stage processing pipeline designed to address these degradation effects without requiring specialised underwater hardware. Starting from a raw degraded image, the system performs histogram-guided colour compensation on the red and blue channels using the green channel as a stable reference, applies Gray World white balancing to neutralise residual colour cast, and then branches into parallel enhancement paths: unsharp-masking for edge and detail recovery, and HSV-domain histogram equalisation for global contrast improvement. The two enhanced streams are subsequently fused through both an averaging strategy and a Principal Component Analysis (PCA) weighted combination. Quality is assessed quantitatively using Mean Squared Error (MSE) and Peak Signal-to-Noise Ratio (PSNR) against reference images. Experimental results demonstrate measurable and visually convincing improvements across a representative set of underwater scenes. The modular design of the pipeline ensures that individual stages can be independently upgraded, making the system readily extensible as algorithmic advances emerge.

Introduction

The paper presents a hybrid image enhancement pipeline for underwater photography, addressing the major problem that underwater images suffer from color distortion, low contrast, scattering, and loss of detail due to light absorption and water turbidity. These issues negatively impact both visual quality and computer vision applications in fields such as marine biology, underwater inspection, and robotics.

The proposed system follows a modular processing workflow rather than a single deep learning model. It includes:

• Color compensation and white balancing to correct blue-green color cast and restore lost red tones
• Contrast enhancement (HSV-based histogram equalization) to improve brightness and visibility
• Sharpness enhancement (unsharp masking) to recover fine details
• Fusion techniques (average and PCA-based) to combine enhanced outputs
• Quality evaluation using MSE and PSNR

The system is implemented in Python using standard libraries (NumPy, PIL, Streamlit) and tested on benchmark datasets such as EUVP, UFO-120, and Sea-Thru. Results show that PCA-based fusion performs better than simple averaging, producing higher PSNR and lower error values.

The study also reviews traditional and deep learning-based approaches, noting that while deep learning methods (CNNs, GANs, transformers) achieve strong performance, they often struggle with generalization across different underwater environments.

Conclusion

This paper presented a modular, multi-stage pipeline for underwater image enhancement, addressing the interrelated degradation effects of wavelength-selective absorption, scattering-induced haze, and loss of contrast and sharpness. The pipeline proceeds from histogram-guided RGB channel compensation through Gray World white balancing, parallel HSV contrast enhancement and unsharp masking, and PCA-weighted image fusion, culminating in quantitative evaluation using MSE and PSNR. Experiments on seven representative underwater images demonstrate that the pipeline consistently reduces MSE and improves PSNR relative to the unenhanced input, with PCA fusion outperforming simple averaging across all test images. The improvements are visually compelling: colour casts are neutralised, natural tones are restored, local contrast is heightened, and fine structural details become legible where previously they were obscured. The modular architecture makes the system readily extensible. Individual stages can be replaced with more sophisticated alternatives — a learned colour correction module, a GAN-based enhancement component, or a transformer-based fusion mechanism — without disrupting the rest of the pipeline. The system therefore serves both as a practical tool for current deployment and as a foundation for future research in this domain.

References

[1] S. K. Nayar and S. G. Narasimhan, “Vision in bad weather,” in Proc. 7th IEEE Int. Conf. on Computer Vision, vol. 2, pp. 820–827, IEEE, 1999. [2] S. G. Narasimhan and S. K. Nayar, “Chromatic framework for vision in bad weather,” in Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), vol. 1, pp. 598–605, IEEE, 2000. [3] R. T. Tan, “Visibility in bad weather from a single image,” in Proc. IEEE Conf. on Computer Vision and Pattern Recognition, pp. 1–8, IEEE, 2008. [4] R. Fattal, “Single image dehazing,” ACM Trans. Graph., vol. 27, no. 3, pp. 1–9, 2008. [5] K. He, J. Sun, and X. Tang, “Single image haze removal using dark channel prior,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 33, no. 12, pp. 2341–2353, 2010. [6] M. J. Islam, Y. Xia, and J. Sattar, “Fast underwater image enhancement for improved visual perception,” IEEE Robot. Autom. Lett., vol. 5, no. 2, pp. 3227–3234, 2020. [7] C. Li et al., “An underwater image enhancement benchmark dataset and beyond,” IEEE Trans. Image Process., vol. 29, pp. 4376–4389, 2020.

Copyright

Copyright © 2026 Mrs. Nalini H C, Meghana Y S, Akshatha K R, Jeevan H R, Likitha D K. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.