Abstract:

Most current image denoising algorithms usually lose the details of an image while removing the image noise, especially when the noise intensity is large, and in some cases, distortion appears. Because the current neural network structure generally tends to have a deep-level design, fusing the shallow features of an image with deep features is difficult. To address these issues, an attention-based two-pass decoder approach for image denoising is proposed. First, a Residual Dense Block (RDB) is designed to increase the depth of the network by improving the U-Net network. The RDB increases the depth of the network, effectively improves the stability of the model, and alleviates the problem of gradient disappearance. Second, a dual decoder structure is designed through multi-scale feature extraction performed by decoders of different scales to strengthen the fusion of shallow and deep features. Third, by introducing an attention mechanism in the decoder, the edge information of the image is captured in a targeted manner to enhance the denoising performance of the model. Experiments prove that, compared with existing common image denoising methods, the proposed method not only effectively removes image noise but also better restores image texture details and has a faster denoising rate. Hence, the proposed method obtains better results in both subjective and objective evaluations.

Key words: deep learning, Convolutional Neural Network(CNN), image denoising, dual decoder, multi-scale

摘要：

目前大多数图像去噪算法在去除图像噪声的同时, 通常会丢失图像的细节信息, 特别是当噪声强度较大时甚至会出现失真。随着当前神经网络结构普遍趋向于深层设计, 导致图像的浅层特征难以与深层特征融合。针对这些问题, 提出一种基于注意力机制的双路解码器图像去噪方法。首先, 设计一种残差密集块(RDB)来对U-Net网络进行改进, 实现网络深度的增加, 有效提升模型的稳定性并缓解梯度消失问题; 其次, 设计一种双路解码器结构, 通过在不同尺度的解码器中进行多尺度特征提取, 加强深浅层特征的融合; 最后, 通过在解码器中引入注意力机制, 有针对性地捕获图像的边缘信息, 增强模型的去噪表现。实验结果表明, 相较于现有常见的图像去噪方法, 所提方法不仅能够有效去除图像噪声, 还能更好地恢复图像纹理细节, 同时具有较快的去噪速度, 在主观和客观评价中均获得了更好的结果。

关键词: 深度学习, 卷积神经网络, 图像去噪, 双路解码器, 多尺度