| 摘 要: 针对养殖水体环境中图像噪声复杂、对虾轮廓模糊的问题，提出了一种基于改进卷积神经网络（CNN）的水下对虾图像去噪方法。该方法由噪声估计子网络和非盲去噪子网络组成，前者采用全卷积结构生成噪声水平图，实现对不同噪声强度的自适应估计；后者基于残差 U-Net 结构，将噪声水平图与原始含噪图像融合输入，从而在有效抑制噪声的同时保持图像细节。为提高去噪质量，设计了结合结构相似性（SSIM）与均方误差（MSE）的复合损失函数，以兼顾像素精度与结构保持。实验结果表明，该算法在公开数据集和自建对虾图像数据集上均取得了优于传统方法（BM3D）和深度学习方法（DnCNN、FFDNet）的性能表现。在 PSNR 和 EPI 指标上，模型分别提升约2dB和0.05，显著改善了图像清晰度与边缘保真度。研究结果表明，该方法能够在复杂水下环境中实现高质量去噪，为智能化水产养殖的图像监测与分析提供可靠的数据支持。 |
| 关键词: 水下图像去噪 卷积神经网络 自适应去噪 图像增强 |
|
中图分类号:
文献标识码:
|
| 基金项目: 国家自然科学基金面上项目(32473216)，宁波市青年科技创新领军人才项目(2023QL004) |
|
| Research on an Underwater Shrimp Image Enhancement Method Based on an Improved Convolutional Neural Network |
|
Xieyufeng1, FanShengli2, Caiweiming2, Wushuangle2, Wanglei3, Wuxin1
|
1.Zhejiang Sci-tech University;2.NingboTech University;3.Zhejiang University
|
| Abstract: To address the problems of complex image noise and blurred shrimp contours in aquaculture water environments, an underwater shrimp image denoising method based on an improved convolutional neural network (CNN) is proposed. The framework consists of a noise estimation subnetwork and a non-blind denoising subnetwork. The former employs a fully convolutional architecture to generate a noise-level map for adaptive estimation of varying noise intensities, while the latter integrates the noise-level map with the original noisy image through a residual U-Net structure to suppress noise effectively while preserving image details. To further improve denoising performance, a composite loss function combining structural similarity (SSIM) and mean squared error (MSE) is designed to balance pixel accuracy and structural fidelity. Experimental results demonstrate that the proposed algorithm outperforms both traditional methods (BM3D) and deep learning-based approaches (DnCNN, FFDNet) on public and self-built shrimp image datasets. The model achieves average improvements of approximately 2 dB in PSNR and 0.05 in EPI, significantly enhancing image clarity and edge fidelity. These results indicate that the proposed method can achieve high-quality denoising in complex underwater environments, providing reliable data support for intelligent aquaculture image monitoring and analysis. |
| Keywords: Underwater image denoising Convolutional neural network Adaptive denoising Image enhancement |
