Deep Learning-Based Multi-Scenario Image Enhancement for Low-Light, Hazy, and Underwater Environments

 Due to its numerous uses in fields including augmented reality, photographic, telemetry, remote sensing and undersea research image enhancement has garnered a lot of attention in recent years. Although deep learning has made significant progress in this field, the majority of cutting edge methods are still domain-specific; concentrating on improving photographs impacted by either low light, haze, or underwater distortions rarely all three inside a single, cohesive system. In order to close this gap, a generic image improvement framework that can handle several difficult visual degradations with single model architecture was proposed and evaluated in this thesis. The proposed framework is built upon an attention based U-Net, which retains the strengths of the original U-Net in extracting multi scale spatial, features and enhances it further through the integration of attention modules. These attention mechanisms empower the network to selectively focus on regions of interest, leading to improved restoration of fine structures, contrast and perceptual details. Importantly, the same model is used across all domains without modification or retraining, thus supporting the central hypothesis that a unified enhancement approach is feasible, efficient and effective. To support the development and evaluation of this approach, a tri-domain paired dataset was constructed by combining representative image pairs from three publicly available and widely recognized datasets: LoLI-Street for low-light images, EUVP for underwater images, and RESIDE for hazy images. Each image pair consisted of a degraded input and its corresponding high quality reference, enabling the model to be trained in a supervised manner. The Adam optimizer and a straightforward but efficient Mean Squared Error (MSE) loss function were utilized to train the model. Despite the absence of complex loss components such as perceptual loss, the model demonstrated high accuracy and visual fidelity. Evaluation was performed using both general purpose metrics - PSNR and SSIM, as well as domain specific metrics like UIQM for underwater enhancement, EPI for low-light exposure quality and FADE for haze density estimation. The results were highly promising. On the low-light test set, the model achieved a PSNR of 36.36 dB and an SSIM of 0.9731, which are considered excellent by both academic and industry standards. In the underwater domain, the model delivered a PSNR of 24.14 dB, SSIM of 0.8461, and a UIQM score of 24.308, reflecting significant improvements in visual clarity and color balance. The overall performance across the unified test set remained consistent, with an average PSNR of 29.13 dB and SSIM of 0.9176. Qualitative analysis further confirmed these findings. Visual comparisons showed that the model successfully brightened dark regions, removed haze and corrected underwater color casts, producing outputs that were perceptually close to the ground truth. Error heat maps and pixel intensity scatter plots provided deeper insights into the behavior of the model, revealing high spatial precision and structural coherence, particularly in the low-light and hazy domains. In conclusion, this thesis has demonstrated that a single, well designed deep learning model, augmented with attention mechanisms, can perform effective enhancement across diverse degradation domains. The unified framework eliminates the need for domain specific networks, reducing complexity and making deployment more scalable in real world applications. This work contributes a significant step toward generalized, efficient and high quality image enhancement under natural and adverse environmental conditions.

[1] S. Anwar and N. Barnes, “Densely residual laplacian super-resolution,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 3, pp. 1192–1204, 2020.  [2] K. G. Lore, A. Akintayo, and S. Sarkar, “Llnet: A deep autoencoder approach to natural low-light image enhancement,” Pattern Recognition, vol. 61, pp. 650–662, 2017. [3] R. C. Gonzalez, Digital image processing. Pearson education india, 2009. [4] Y. Zhang, X. Di, B. Zhang, and C. Wang, “Self-supervised image enhancement network: Training with low light images only,” arXiv preprint arXiv:2002.11300, 2020. [5] D. Akkaynak and T. Treibitz, “Sea-thru: A method for removing water from underwater images,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 1682–1691. [6] K. He, J. Sun, and X. Tang, “Single image haze removal using dark channel prior,” IEEE transactions on pattern analysis and machine intelligence, vol. 33, no. 12, pp. 2341–2353, 2010. [7] Z. Wang, J. Chen, and S. C. H. Hoi, “Deep Learning for Image SuperResolution: A survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 43, no. 10, pp. 3365–3387, 3 2020. [Online]. Available: https://doi.org/10.1109/tpami.2020.2982166 [8] O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical image computing and computerassisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. Springer, 2015, pp. 234–241. [9] O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Misawa, K. Mori, S. McDonagh, N. Y. Hammerla, B. Kainz et al., “Attention u-net: Learning where to look for the pancreas,” arXiv preprint arXiv:1804.03999, 2018.  [10] C. Wei, W. Wang, W. Yang, and J. Liu, “Deep retinex decomposition for low-light enhancement,” arXiv preprint arXiv:1808.04560, 2018.