Binarizing Super-Resolution Neural Network Without Batch Normalization.

Xunchao Li,Fei Chao

DOI: https://doi.org/10.1007/978-981-99-8549-4_6

2024-01-01

Abstract:In this paper, our objective is to propose a model binarization method aimed at addressing the challenges posed by over-parameterized super-resolution (SR) models. Our analysis reveals that binary SR models experience significant performance degradation, primarily attributed to their sensitivity towards weight/activation distributions, particularly when devoid of Batch Normalization (BN) layers. Consequently, we undertake the following endeavors in this study: First, we conduct a comprehensive analysis to examine the impact of BN layers on SR models based on Binary Neural Networks (BNNs). Second, we propose an asymmetric binarizer that can be reparameterized to adaptively adjust the transition point for activation binarization. Third, we introduce a progressive gradient estimator that modifies weight smoothness and controls weight flipping to stabilize the training procedure in the absence of BN layers. Through extensive experiments, we demonstrate that our proposed method exhibits significant performance improvements. For instance, when binarizing EDSR and scaling up input images by a factor of x4, our approach achieves a PSNR decrease of less than 0.4dB on the Urban100 benchmark.