Compress Clean Signal from Noisy Raw Image: A Self-Supervised Approach
Authors: Zhihao Li, Yufei Wang, Alex Kot, Bihan Wen
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that our method surpasses existing compression techniques, achieving a more advantageous rate-distortion balance with improvements ranging from +2 to +10d B and yielding a bit saving of 2 to 50 times. We curate a full-day dataset of raw images with calibrated noise parameters and reference images to evaluate the performance of models under a wide range of input signalnoise ratios. In this section, we present the Rate-Distortion (RD) curves across various SNR levels for the SID, ELD, and FDRIC datasets. We present several ablation studies to validate our method. |
| Researcher Affiliation | Academia | 1Department of EEE, Nanyang Technology University, Singapore. Correspondence to: Bihan Wen <bihan.wen@ntu.edu.sg>. |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available at https: //lizhihao6.github.io/Cleans. |
| Open Datasets | Yes | Existing raw image denoising datasets mainly focus on low-light or nighttime scenes, e.g., SID (Chen et al., 2018) captured under around 5 lux conditions and ELD (Wei et al., 2020) is even below 0.3 lux as shown in Fig. 4a. To address this limitation, we develop a full-day raw image compression (FDRIC) dataset that covers a wide range of SNR, ensuring comprehensive training and evaluation. |
| Dataset Splits | No | The paper states 'We follow the train-test set split for the SID (2018) Sony A7S2 subset as outlined in PMN (2022).' and 'For our FDRIC dataset, we crop these images into 512 512 patches for both training and testing.', but does not explicitly detail the validation dataset split or provide specific percentages/counts for it. |
| Hardware Specification | Yes | All the models are trained within Compress AI Py Torch framework (2020) using a single NVIDIA RTX 3090 GPU. |
| Software Dependencies | No | The paper mentions 'All the models are trained within Compress AI Py Torch framework (2020)', but it does not specify version numbers for other key software dependencies or libraries (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | Adam optimizer with an initial learning rate of 1e-4 and a batch size of 6 is used, spanning 200,000 iterations with a learning rate decay to 1e-5 after 150,000 iterations. These hyperparameters remain consistent across all datasets, and we apply grad norm clipping for training stability as in RIC. For the hyperparameters λD, λP , and λcov in Eq. (23) and Eq. (25), the range of λD is set between 0.0018 and 0.18. The minimum value of λP is 0.05, and the maximum at 5, increasing at the same rate as λD. λcov is consistently maintained at 0.2. |