Rethinking Transformer-Based Blind-Spot Network for Self-Supervised Image Denoising
Authors: Junyi Li, Zhilu Zhang, Wangmeng Zuo
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on real-world image denoising datasets show that TBSN largely extends the receptive field and exhibits favorable performance against state-of-the-art SSID methods. ... Experiments Implementation Details Datasets. We conduct experiments on two wildly used real-world image denoising datasets... |
| Researcher Affiliation | Academia | Harbin Institute of Technology, Harbin, China EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper includes a Python code snippet to visualize the receptive field, stating "The pseudo code in Py Torch format is as follows...", but it is not labeled as a core algorithm or pseudocode for the main methodology of the paper. |
| Open Source Code | Yes | Code https://github.com/nagejacob/TBSN |
| Open Datasets | Yes | We conduct experiments on two wildly used realworld image denoising datasets, i.e., SIDD (Abdelhamed, Lin, and Brown 2018) and DND (Plotz and Roth 2017). |
| Dataset Splits | Yes | SIDD dataset ... It contains 320 training images, 1280 validation patches and 1280 benchmark patches, respectively. We train our networks on the noisy images of train split, and test on the benchmark split. ... DND is a benchmark dataset ... It contains 50 pairs for test only. We train and test our networks on the test images in a fully self-supervised manner. |
| Hardware Specification | Yes | All the experiments are conducted on Py Torch framework and Nvidia RTX2080Ti GPUs. |
| Software Dependencies | No | The paper mentions using "Py Torch framework" but does not specify a version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | The batch size and patch size are set to 4 and 128 128, respectively. We adopt ℓ1 loss and Adam W (Loshchilov and Hutter 2018) optimizer to train the network. The learning rate is initially set to 3 10 4, and is decreased by 10 every 40k iterations with total 100k training iterations. |