This is official PyTorch implementation of "Enhancing Self-Supervised Image Denoising with Asymmetric Mask Blind-Spot Network."
Image denoising is a fundamental task in image processing and computer vision. Traditional supervised methods heavily rely on large sets of noisy-clean image pairs, which are impractical to obtain for real-world noisy images. Self-supervised denoising methods offer a viable alternative but often struggle with spatial correlation in noise. In this paper, we introduce the Asymmetric Mask Blind-Spot Network (AM-BSN), designed to disrupt spatial correlations of large-scale noise in real-world images. Our network features a dual-branch architecture: a local branch employing a 3x3 central mask convolution for fine detail recovery, and a global branch utilizing a 5x5 'X'-shaped mask convolution and dilated convolutions for global structure reconstruction. Experimental results on real-world datasets demonstrate that AM-BSN outperforms state-of-the-art methods, achieving a PSNR of 37.90 dB and an SSIM of 0.885 on the SIDD benchmark. This research advances self-supervised denoising techniques, providing a practical solution for real-world applications.
| Models | SIDD Validation | Parameters |
|---|---|---|
| AP-BSN | 35.91/0.870 | 3.7M |
| MM-BSN | 37.38/0.882 | 5.3M |
| AM-BSN | 37.54/0.884 | 3.8M |
Our experiments are done with:
- Python 3.10.13
- PyTorch 2.1.1+cu118
- numpy 1.26.3
- opencv 4.10.0.84
- scikit-image 0.24.0
The data used in the training of our method is consistent with that of the APBSN. For the specific method of obtaining training data, please refer to AP-BSN.
To test noisy images with pre-trained AM-BSN in gpu:0
Use png2csv.py to convert the denoised images from the SIDD benchmark dataset into the CSV format required for the Kaggle competition.
python test.py -c SIDD -g 0 --pretrained ./ckpt/AMBSN.pth --td [your noisy images dir]
As soon as our paper is accepted, we will upload the training code immediately.
Part of our codes are adapted from AP-BSN and MMBSN. We are expressing gratitude for their work sharing.