Area-Efficient FPGA Implementation of Minimalistic Convolutional Neural Network Using Residue Number System

Convolutional Neural Networks (CNN) is the promising tool for solving task of image recognition in computer vision systems. However, the most known implementation of CNNs require a significant amount of memory for storing weights in training and work. To reduce the resource costs of CNN implementation we propose the architecture that separated on hardware and software parts for performance optimization. Also we propose to use Residue Number System (RNS) arithmetic in the hardware part which implements the convolutional layer of CNN. Software simulation using Matlab 2017b shows that CNN with a minimum number of layers can be quickly and successfully trained. Hardware simulation using FPGA Kintex7 xc7k70tfbg484-2 demonstrates that using RNS in convolutional layer of CNN allows to reduce hardware costs by 32% compared with the traditional approach based on the binary number system.