Quantization-friendly super-resolution: Unveiling the benefits of activation normalization

Abstract

Super-resolution (SR) has achieved remarkable progress with deep neural networks, but the substantial memory and computational demands of SR networks limit their use in resource-constrained environments. To address these challenges, various quantization methods have been developed, focusing on managing the diverse and asymmetric activation distributions in SR networks. This focus is crucial, as most SR networks exclude batch normalization (BN) due to concerns about image quality degradation from limited activation range flexibility. However, this decision is made in the context of full-precision SR networks, leaving BN’s impact on quantized SR networks uncertain. This paper revisits BN’s role in quantized SR networks, presenting a detailed performance analysis of multiple quantized SR models with and without BN. Experimental results show that including BN in quantized SR networks enhances performance and simplifies network design through minor yet significant structural adjustments. These findings challenge conventional assumptions and offer new insights for SR network optimization.