A Simple and General Binarization Method for Image Restoration Neural Networks

Mengxue Wang, Yue Zhang, Xiaodong Zhang, Run Min
Proceedings of the 15th Asian Conference on Machine Learning, PMLR 222:1433-1448, 2024.

Abstract

With the advancement of deep learning techniques, image restoration (IR) performance has improved significantly. However, these techniques often come with high computational costs, which pose challenges in meeting the processing latency requirements of resource-constrained hardware in edge computer vision systems. To address this issue, we propose a simple binarization technique and an efficient training strategy called Gentle Approximation Method (GAM) to extend the application of binary neural networks (BNNs) to various IR tasks, including low-light image enhancement, deraining, denoising, and super-resolution. Our results demonstrate the effectiveness of our method in binarizing full-precision deep neural networks. By binarizing these networks, we achieve a significant reduction in computational and memory demands while maintaining satisfactory performance. For instance, in the denoising task, the FLOPs can be reduced to only 3% of the original network while preserving most of the performance.

Cite this Paper

BibTeX
@InProceedings{pmlr-v222-wang24b, title = {A Simple and General Binarization Method for Image Restoration Neural Networks}, author = {Wang, Mengxue and Zhang, Yue and Zhang, Xiaodong and Min, Run}, booktitle = {Proceedings of the 15th Asian Conference on Machine Learning}, pages = {1433--1448}, year = {2024}, editor = {Yanıkoğlu, Berrin and Buntine, Wray}, volume = {222}, series = {Proceedings of Machine Learning Research}, month = {11--14 Nov}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v222/wang24b/wang24b.pdf}, url = {https://proceedings.mlr.press/v222/wang24b.html}, abstract = {With the advancement of deep learning techniques, image restoration (IR) performance has improved significantly. However, these techniques often come with high computational costs, which pose challenges in meeting the processing latency requirements of resource-constrained hardware in edge computer vision systems. To address this issue, we propose a simple binarization technique and an efficient training strategy called Gentle Approximation Method (GAM) to extend the application of binary neural networks (BNNs) to various IR tasks, including low-light image enhancement, deraining, denoising, and super-resolution. Our results demonstrate the effectiveness of our method in binarizing full-precision deep neural networks. By binarizing these networks, we achieve a significant reduction in computational and memory demands while maintaining satisfactory performance. For instance, in the denoising task, the FLOPs can be reduced to only 3% of the original network while preserving most of the performance.} }
Endnote
%0 Conference Paper %T A Simple and General Binarization Method for Image Restoration Neural Networks %A Mengxue Wang %A Yue Zhang %A Xiaodong Zhang %A Run Min %B Proceedings of the 15th Asian Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2024 %E Berrin Yanıkoğlu %E Wray Buntine %F pmlr-v222-wang24b %I PMLR %P 1433--1448 %U https://proceedings.mlr.press/v222/wang24b.html %V 222 %X With the advancement of deep learning techniques, image restoration (IR) performance has improved significantly. However, these techniques often come with high computational costs, which pose challenges in meeting the processing latency requirements of resource-constrained hardware in edge computer vision systems. To address this issue, we propose a simple binarization technique and an efficient training strategy called Gentle Approximation Method (GAM) to extend the application of binary neural networks (BNNs) to various IR tasks, including low-light image enhancement, deraining, denoising, and super-resolution. Our results demonstrate the effectiveness of our method in binarizing full-precision deep neural networks. By binarizing these networks, we achieve a significant reduction in computational and memory demands while maintaining satisfactory performance. For instance, in the denoising task, the FLOPs can be reduced to only 3% of the original network while preserving most of the performance.
APA
Wang, M., Zhang, Y., Zhang, X. & Min, R.. (2024). A Simple and General Binarization Method for Image Restoration Neural Networks. Proceedings of the 15th Asian Conference on Machine Learning, in Proceedings of Machine Learning Research 222:1433-1448 Available from https://proceedings.mlr.press/v222/wang24b.html.

Related Material