Normalization-Equivariant Neural Networks with Application to Image Denoising
Authors: Sébastien Herbreteau, Emmanuel Moebel, Charles Kervrann
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results in image denoising show that normalization-equivariant neural networks, in addition to their better conditioning, also provide much better generalization across noise levels. |
| Researcher Affiliation | Academia | Sébastien Herbreteau Emmanuel Moebel Charles Kervrann Centre Inria de l Université de Rennes, France {sebastien.herbreteau, emmanuel.moebel, charles.kervrann}@inria.fr |
| Pseudocode | No | The paper describes the proposed architectural modifications (affine convolutions and channel-wise sort pooling) through descriptive text and diagrams (Figure 2), but it does not provide any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code and pre-trained models can be downloaded here: https://github.com/sherbret/normalization_equivariant_nn/. |
| Open Datasets | Yes | We use the same large training set as in [46] for all the models and all the experiments, composed of 8, 694 images, including 400 images from the Berkeley Segmentation Dataset BSD400 [32], 4, 744 images from the Waterloo Exploration Database [28], 900 images from the DIV2K dataset [1], and 2, 750 images from the Flickr2K dataset [25]. |
| Dataset Splits | Yes | The dataset BSD32 [32], composed of the 32 images, is used as validation set to control training and select the best model at the end. |
| Hardware Specification | Yes | GPU: Quadro RTX 6000, CPU: 2,3 GHz Intel Core i7. |
| Software Dependencies | No | The paper mentions 'Python' and 'Py Torch library [35]' but does not specify version numbers for these software components. |
| Experiment Setup | Yes | The training parameters, specific to each model and its variants, are guided by the instructions of the original papers [46, 47], to the extent possible, and are summarized in Table 3. Note that each training iteration consists in a gradient pass on a batch composed of patches randomly cropped from training images. Normalization-equivariant variants need a longer training and always use a constant learning rate (speed improvements are however certainly possible by adapting the learning rate throughout optimization, but we did not investigated much about it). Furthermore, contrary to [46] where the ℓ1 loss function is recommended to achieve better performance, supposedly due to its outlier robustness properties, we obtained slightly better results with the usual mean squared error (MSE) loss when dealing with normalization-equivariant networks. |