Image Quality Improvement of Surveillance Camera Images by Learning-based Denoising Method Utilizing Noise2Noise
Image Quality Improvement of Surveillance Camera Images by Learning-based Denoising Method Utilizing Noise2Noise
复制标题
DOI：
10.1145/3576938.3576943
复制
发表时间：
2022-11
期刊：
影响因子：
--
通讯作者：
Akira Kuchida;T. Goto
中科院分区：
文献类型：
--
作者：
Akira Kuchida;T. Goto
近年来，监控摄像头的安装数量有所增加。监控摄像机需要能够在低曝光等恶劣拍摄条件下捕捉图像。然而，在这种环境下，捕获的图像可能会产生噪声。虽然对图像去噪的研究很多，但大多只针对合成噪声，如高斯噪声或真实图像噪声，如SIDD数据集，对捕获的图像没有足够的性能。在本文中，我们研究了使用Noise2Noise构建一个有效的真实图像噪声CNN模型。此外，当数据量较小时，Noise2Noise与正常学习相比存在性能明显下降的问题。因此，我们提出了一种学习方法，通过使用开放数据集（如SIDD）进行预训练，然后使用Noise2Noise进行重新训练，即使在数据很小的情况下也可以构建具有良好性能的模型。
In recent years, the number of surveillance cameras installed has increased. Surveillance cameras need to be able to capture images even under poor shooting conditions such as low exposure. However, noise may be generated in the captured images under such environments. Although there have been many studies on image denoising, most of them target only synthetic noise such as Gaussian noise or real image noise such as the SIDD dataset and have not demonstrated sufficient performance for captured images. In this paper, we investigate the construction of an effective CNN model for real image noise using Noise2Noise. In addition, Noise2Noise has the problem of significantly degraded performance compared to normal learning when data is small. Therefore, we propose a learning method that can build models with good performance even when data is small, by pre-training with an open dataset such as SIDD and then re-training with Noise2Noise.