Towards Boosting Channel Attention for Real Image Denoising: Sub-band Pyramid Attention
Towards Boosting Channel Attention for Real Image Denoising: Sub-band Pyramid Attention
复制标题
DOI：
10.1007/978-3-030-87361-5_25
复制
发表时间：
2021
期刊：
影响因子：
--
通讯作者：
Huayu Li;Haiyu Wu;Xiwen Chen;Hao Wang;A. Razi
中科院分区：
文献类型：
--
作者：
Huayu Li;Haiyu Wu;Xiwen Chen;Hao Wang;A. Razi
卷积层平等对待通道特征，没有优先级。当卷积神经网络（CNN）用于具有未知噪声分布的真实应用中的图像去噪时，特别是具有可学习模式的结构化噪声，建模信息特征可以大大提高去噪性能。真实世界图像去噪任务中的通道注意力利用特征通道之间的依赖性;因此，它们可以被视为频域滤波机制。现有的信道注意模块通常使用全局静态作为描述符来学习信道间相关性。这些方法被认为在学习用于在频率水平上重新缩放信道的代表性系数方面效率低下。提出了一种基于小波变换的子带金字塔注意力（SPA）模型，以更细粒度的方式重新校准提取的特征的频率分量。从某种意义上说，我们的方法将传统的频域滤波方法与深度学习架构相结合，以实现更高的性能记录。实验结果表明，人工神经网络配备了建议的注意力模块大大提高了基准天真通道注意力块。更具体地说，我们获得了3.97 dB的增益相比，最好的传统算法，BM 3D和1.87 dB至0.18 dB的增益在基于DL的方法在去噪性能。此外，我们的研究结果表明金字塔的水平如何影响的SPA块的性能，并表现出良好的推广能力的SPA块。
Convolutional layers treat the Channel features equally with no prioritization. When Convolutional Neural Networks (CNNs) are used for image denoising in real-world applications with unknown noise distributions, particularly structured noise with learnable patterns, modeling informative features can substantially boost the denoising performance. Channel attentions in real-world image denoising tasks exploit dependencies between the feature channels; therefore, they can be viewed as a frequency-domain filtering mechanism. Existing channel attention modules typically use global statics as descriptors to learn inter-channel correlations. These methods deem inefficient in learning representative coefficients for re-scaling the channels at frequency level. This paper proposes a novel Sub-band Pyramid Attention (SPA) model based on wavelet transform to recalibrate the extracted features’ frequency components in a more fine-grained fashion. Our method, in one sense, integrates the conventional frequency-domain filtering methods with deep learning architectures to achieve higher performance records. Experimental results show that ANNs equipped with the proposed attention module substantially improves upon the benchmark naive channel attention blocks. More specifically, we obtained a 3.97 dB gain compared to the best traditional algorithm, BM3D and a 1.87 dB to 0.18 dB gain over the DL-based methods in terms of denoising performance. Furthermore, our results show how the pyramid level affects the performance of the SPA blocks and exhibits favorable generalization capability for the SPA blocks.