Abstract:

Deep neural networks have a strong denoising ability for real images and can learn complex nonlinear mapping relationships between noisy and clean images. However, excessive convolution operations result in increased computational costs and occupy a large amount of memory, limiting the application of denoising techniques in low computing power devices. Existing denoising algorithms are prone to damaging detail information, and restoring images may suffer from problems such as excessively smooth edges, missing textures, and residual noise. To address these issues, construct a Multi-level Residual Information Distillation Block (MRIDB) is constructed in this study. By segmenting feature channels, retaining some features for subsequent multi-level fusion, and further extracting refined feature information through depth extraction units. It introduces a Contrast-aware Channel Attention (CCA) mechanism to assign weights to features of different channels. It uses multi-level skip connections to fully integrate contextual information extracted from different stages. It builds a lightweight Multi-level Residual Information Distillation Network (MIRDN) using a low inter- block complexity encoding-decoding structure. The encoding part is a noisy image feature extraction module, and whereas the decoding part is a clean image restoration module. In order to accelerate the training speed, a progressive training method with mixed image sizes is adopted. The experimental results show demonstrate that the Peak Signal-to-Noise Ratios (PSNR) of the proposed method on SSID and DND real image datasets are 39.43 dB and 39.49 dB, respectively. These are improvements in the ranges of 0.17-15.77 dB and 0.02-7.06 dB, respectively, compared with other networks, while and the model parameter count is only 6.92×10⁶. The proposed model has fewer parameters while improving denoising performance.

Key words: image restoration, real image denoising, Multi-level Residual Information Distillation Block(MRIDB), Deep Extraction Module(DEM), Contrast-aware Channel Attention(CCA)

摘要：

深度神经网络对真实图像有较强的去噪能力，可以学习含噪图像和干净图像之间复杂的非线性映射关系。然而，过多的卷积操作导致计算成本增加并占据大量内存，限制了去噪技术在低运算能力设备中的应用，现有去噪算法容易损坏细节信息，恢复图像存在边缘过度平滑、纹理缺失、含有残留噪声等问题。针对这些问题，构造一种多级残差信息蒸馏模块。通过对特征通道进行分割，保留部分特征用于后续多级融合，并进一步通过深度提取单元提取细化后的特征信息；引入对比度感知通道注意力机制对不同通道的特征分配权重；使用多级跳跃连接充分融合不同阶段提取到的上下文信息。构建1个轻量级的多级残差信息蒸馏网络，采用块间复杂度低的编码-解码结构，编码部分为含噪图像特征提取模块，解码部分为干净图像恢复模块。为了加快训练速度，采用混合图像尺寸的渐进式训练方法。实验结果表明，该方法在SSID和DND真实图像数据集上的峰值信噪比分别为39.43 dB和39.49 dB，与其他网络相比提升了0.17~15.77 dB和0.02~7.06 dB，而模型参数量仅为6.92×10⁶，所提模型在提高去噪性能的同时具有较少的参数量。

关键词: 图像复原, 真实图像去噪, 多级残差信息蒸馏模块, 深度提取模块, 对比度感知通道注意力