Back to articles
Volume: 34 | Article ID: ISS-183
On quantization of convolutional neural networks for image restoration
  DOI :  10.2352/EI.2022.34.7.ISS-183  Published OnlineJanuary 2022

Recently, commercial vision sensors hit the mobile market. To achieve that, computer vision networks had to be quantized. However, this topic was not studied well for Image Signal processor (ISP) challenging image restoration tasks, being crucially important for hardware implementation, as well as for deployment on hardware accelerators, e.g. Neural Processors Units (NPU). In this paper, we studied the effect of the quantization of deep learning network on image quality. We tried various quantization on raw RGBW image demosaicing. Experimental results show that 12 bit weight quantization can sustain image quality at the similar level with floating-point network. 10 bit quantized network shows slight degradation in objective image quality and mild visual artifacts. If network weight’s bit-depth can be significantly reduced for computer vision tasks, our finding shows that it is not true for raw image restoration tasks: at least 10 bit weights are required to provide sufficient image quality. However, we can save some memory on feature maps bit-depth. We can conclude that network bit depth is critical for raw image restoration.

Subject Areas :
Views 66
Downloads 21
 articleview.views 66
 articleview.downloads 21
  Cite this article 

Youngil Seo, Irina Kim, Jeongguk Lee, Wooseok Choi, Seongwook Song, "On quantization of convolutional neural networks for image restorationin Proc. IS&T Int’l. Symp. on Electronic Imaging: Imaging Sensors and Systems,  2022,  pp 183-1 - 183-5,

 Copy citation
  Copyright statement 
Copyright © Society for Imaging Science and Technology 2022
Electronic Imaging
Society for Imaging Science and Technology
IS&T 7003 Kilworth Lane, Springfield, VA 22151 USA