基于双通道残差网络的水下图像去噪研究

Research on Underwater Image Denoising Based on Dual-Channels Residual Network

水下图像去噪作为探索海底世界有效的辅助手段之一,备受研究人员的关注。传统的滤波方法存在去噪过程中容易损坏图像的细节,带有明显的噪声残留,且根据不同的噪声类型设计相应滤波器的问题,在深度卷积神经网络的基础上,提出一种改进的双通道残差卷积神经网络模型,用于去除水下图像的噪声。该模型由局部残差网(LRN)、全局稀疏网(GSN)和特征处理块(FPB)构成。通过双通道LRN和GSN并行提取水下图像的多层次局部噪声特征和全局噪声特征,利用FPB中的通道连接融合LRN和GSN提取的噪声特征,并使用其卷积层增强水下图像噪声信息。在此基础上,使用均方误差和平均绝对误差双损失函数优化网络参数,利用卷积层重构水下图像。实验结果表明,相比BM3D、IRCNN、DnCNN等方法,该方法的平均峰值信噪比提高0.02~3.52 dB,在有效去除各种水平的随机噪声同时能重构清晰的水下图像。

As an effective underwater exploration tool,underwater image denoising has received increasing attention and research in academia.To address the vulnerability of traditional filtering methods to image degradation in the denoising process,with obvious noise residues,as well as facilitate the design of filters tailored to specific types of noise,an improved dual-channels residual Convolutional Neural Network(CNN)model is proposed herein on the basis of a deep CNN to eliminate noise from underwater images.The proposed model consists of a Local Residual Network(LRN),Global Sparse Network(GSN),and Feature Processing Block(FPB).The multi-level local noise and global noise features of underwater images are extracted via the dual-channels LRN and GSN in parallel.The channel connection in FPB is then used to fuse the extracted noise features from LRN and GSN,and the convolution layer is employed to enhance the noise information.On this basis,two double loss functions-Mean Square Error(MSE)and Mean Absolute Error(MAE)are used to optimize the network parameters,and the convolution layer is used to reconstruct the underwater images.The experimental results show that compared with that of BM3D,IRCNN,DnCNN,and other methods,the average Peak Signal-to-Noise Ratio(PSNR)of the proposed method increases by 0.02~3.52 dB.Thus,the proposed method successfully removes various levels of random noise and reconstructs clear underwater images.