【作者】 李庆辉； 阎威武； 祝孟根； 韩永强；

【Author】 Qinghui Li;Weiwu Yan;Menggen Zhu;Yongqiang Han;School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University;

【机构】 上海交通大学电子信息与电气工程学院；

【摘要】 相比于基于大规模图像对的深度监督学习模型,盲点网络仅需要噪声图像就可以完成训练,训练数据获取成本低。面对当前盲点网络结构设计不合理导致网络训练困难、准确率较低的问题,本文设计了一种基于噪声先验的双通路降噪模型,局部通路和全局通路两条通路有效保证多尺度视野,保留CNN模型局部特征提取能力的同时,通过移动窗口注意力机制引入长距离依赖的全局特征,基于噪声先验假设设计的多空心卷积层能有效地排除噪声相关性对于盲点网络的影响,通道-像素混合层的设计能高效避免盲点网络的退化,以上设计共同完善盲点网络结构,最终稳定训练过程、提高模型准确性。实验结果证明,基于噪声先验的双通路降噪模型仅需要噪声图像即可达到不弱于一些监督网络的降噪效果,且相比于当前的无监督方法也具有显著的优势。更多还原

【Abstract】 Compared to deep supervised learning models based on large-scale image pairs, the blind spot network only requires noisy images for training, resulting in low training data acquisition costs. However, the blind spot network faces challenges in training and lower accuracy. To address this, this paper proposes a dual-path denoising model based on noise priors. The local path and global path effectively ensure a multi-scale view, retaining the local feature extraction capability of CNN models.Meanwhile, by introducing a moving window attention mechanism to incorporate long-range dependencies of global features,the model is designed to alleviate the training difficulties and improve accuracy. The multi-hollow convolution layers designed based on the noise prior hypothesis effectively eliminate the influence of noise correlation on the blind spot network. The channel-pixel fusion layers efficiently prevent the degradation of the blind spot network. These designs collectively stabilize the training process and enhance model accuracy. Experimental results demonstrate that the dual-path denoising model based on noise priors achieves denoising effects not inferior to some supervised networks using only noisy images, and it also exhibits significant advantages over current unsupervised methods.更多还原