Zhaoming Kong ([email protected]), Fangxi Deng, Haomin Zhuang, Jun Yu, Lifang He and Xiaowei Yang
If you do not have access to Google services, please contact me.
In this project, we intend to collect and compare various denoising methods to investigate their effectiveness, efficiency, applicability and generalization ability with both synthetic and real-world experiments. Datasets, code and results are made publicly available and will be continuously updated. To gain access to the dataset and the code, please send me an email ([email protected]) with information including name, position and usage.
The interest in the realm of denoising grows consistently with a large quantity of approaches, which may be roughly divided into two categories, namely traditional denoisers and DNN methods, depending on whether neural network architectures are utilized.
For traditional denoisers, learning and denoising are usually accomplished only with the noisy image by leveraging the NLSS property. To achieve this goal, the most popular and successful framework is attributed to BM3D, which mainly follows three consecutive stages: grouping, collaborative filtering and aggregation. The flowchart of this effective three-stage paradigm is illustrated in the following Figure.
Since the birth of BM3D, there is no shortage of extensions originating from different disciplines. Some representative traditional denoisers are summarized in the following Table.
Representative traditional denoisers (click here)
The most recent development of image processing stems largely from the applications of deep learning techniques, which demonstrate outstanding performance in a wide variety of tasks. Image denoising is not an exception. From the early plain networks to recently proposed generative and diffusion models, numerous net- work architectures and frameworks have been developed with different training strategies, including supervised, self-supervised and unsupervised learning. The following figure illustrates a simple DNN denoising framework with three convolutional layers.
Since the birth of BM3D, there is no shortage of extensions originating from different disciplines. Some representative traditional denoisers are summarized in the following Table.
Related DNN methods (click here)
In addition to existing datasets, in this paper, we introduce the real-world indoor-outdoor color image (IOCI) and video (IOCV) datasets for benchmarking. We use 13 different camera devices to capture images in both outdoor and indoor environments. Instead of using predefined camera settings such as ISO, shutter speed and aperture, we mostly resort to the cameras’ auto mode. In uncontrollable and certain dark environments, we use short exposures and increase ISO values to produce images with high noise levels. Our datasets will be continuously updated once new camera devices are available.
- Image denoising (sRGB space).
- Video denoising (sRGB space).
- MSI/HSI denoising.
- 3D MRI denoising.
