Convolutional neural networks are a promising tool for solving the problem of pattern recognition. Most well-known convolutional neural networks implementations require a significant amount of memory to store weights in the process of learning and working. We propose a convolutional neural network architecture in which the neural network is divided into hardware and software parts to increase performance and reduce the cost of implementation resources. We also propose to use the residue number system (RNS) in the hardware part to implement the convolutional layer of the neural network. Software simulations using Matlab 2018b showed that convolutional neural network with a minimum number of layers can be quickly and successfully trained. The hardware implementation of the convolution layer shows that the use of RNS allows to reduce the hardware costs on 7.86%–۳۷٫۷۸% compared to the two’s complement implementation. The use of the proposed heterogeneous implementation reduces the average time of image recognition by 41.17%.

Introduction

The modern development of science and technology implies the widespread introduction of data mining methods. The range of tasks requiring the use of artificial intelligence methods in the field of image processing is constantly expanding: personal identification [5], scene recognition [6], information processing from external sensors in unmanned land and aircraft vehicles [28], medical diagnostics [23], and so on. The use of intelligent methods based on artificial neural networks is a promising tool for solving the problem of image recognition [35]. The idea of using artificial neural networks for processing visual information was proposed in [15] to solve the problem of automating the recognition of handwritten numbers. The architecture proposed in this paper was called the Convolutional Neural Network (CNN). The combination of convolutional layers, realizing the extraction of visual signs, and a multilayer perceptron, realizing the recognition operation according to the convolution results, is its main feature. The evolution of this scientific idea and the development of computer technology have led to the fact that at present the theory of CNN and its practical application methods are developing along the path of an extensive increase in the number of layers of CNN, which leads to the high computational complexity of the implementation of such systems. For example, the network architecture [14], which showed the best ImageNet image recognition result in 2010, contains about 650 thousand neurons, 60 million tunable parameters and requires 27 GB of disk space for training. The paper [30] presents the development of Google, which showed the best image recognition result of ImageNet base in 2014. This CNN performs more than one and a half billion computational operations for image recognition, which motivated Google to develop a special tensor processing unit to optimize the operation of this CNN [13]. Summarizing, modern CNN architectures are very resource intensive, which limits the possibilities for their wide practical application. A unified approach to reducing resource costs in the implementation of CNN in practice is not currently available.