Sarala A, Vinoth Kumar C
Department of Electronics and Communication Engineering, Kings Engineering College, Chennai, 602117, India.
Department of Electronics and Communication Engineering, Sri Sivasubramaniya Nadar College of Engineering, Chennai, 603110, India.
Sci Rep. 2025 Aug 20;15(1):30480. doi: 10.1038/s41598-025-96832-8.
Underwater image enhancement poses unique challenges due to poor visibility, color distortion, and haze caused by light absorption and scattering in water. In this paper, we propose an ensemble model, Ensemble Pyramid-based Convolutional Neural Network and Deep Channel Prior Dehazing Network (EPCNN-DCPDN), which combines Pyramid-based Convolutional Neural Networks (CNNs) and the Deep Channel Prior Dehazing Network (DCPDN) to address these challenges. The model operates in two ways: sequentially, by first applying DCPDN for haze removal followed by Pyramid-based CNNs for multi-scale feature refinement, or in parallel, with outputs from both models fused using a weighted average or learned fusion mechanism. We evaluated the proposed model on multiple underwater datasets and compared its performance against nine state-of-the-art models, including CLAHE, FUnIE-GAN, WaterGAN, and Haze-Line Prior Model. The EPCNN-DCPDN model achieved superior results with a PSNR of 28.34 dB, SSIM of 0.902, and UIQM of 3.56. It also demonstrated outstanding accuracy in challenging underwater conditions, with an accuracy of 97.92% on shallow, deep, and low-light underwater datasets, outperforming existing models such as WaterGAN and Haze-Line Prior Model. The results highlight the effectiveness of the proposed model in restoring color, contrast, and fine details in underwater images. The model's ability to handle a wide range of underwater conditions makes it an ideal solution for applications in underwater exploration, marine research, and object detection.
由于水中光吸收和散射导致能见度差、颜色失真和雾霾，水下图像增强面临着独特的挑战。在本文中，我们提出了一种集成模型，即基于金字塔的卷积神经网络和深度通道先验去雾网络（EPCNN-DCPDN），它结合了基于金字塔的卷积神经网络（CNN）和深度通道先验去雾网络（DCPDN）来应对这些挑战。该模型有两种运行方式：顺序运行，即先应用DCPDN去除雾霾，然后应用基于金字塔的CNN进行多尺度特征细化；或者并行运行，使用加权平均或学习融合机制融合两个模型的输出。我们在多个水下数据集上评估了所提出的模型，并将其性能与包括CLAHE、FUnIE-GAN、WaterGAN和雾霾线先验模型在内的九个先进模型进行了比较。EPCNN-DCPDN模型取得了优异的结果，PSNR为28.34 dB，SSIM为0.902，UIQM为3.56。它在具有挑战性的水下条件下也表现出了出色的准确性，在浅、深和低光水下数据集上的准确率为97.92%，优于WaterGAN和雾霾线先验模型等现有模型。结果突出了所提出模型在恢复水下图像颜色、对比度和精细细节方面的有效性。该模型处理各种水下条件的能力使其成为水下探索、海洋研究和目标检测应用的理想解决方案。
