FPGA-Based Convolutional Neural Network for Classifying Image Blocks

This article deals with the problem of developing a neural network for processing visual data using programmable logic integrated circuits (ICs). The article also gives an overview of important aspects when choosing the type of neural network architecture for its synthesis into an electrical circuit. Particular attention is paid to the amount of hardware resources used to solve the problem, as well as the speed of the developed system. The convolutional neural network was chosen for this purpose. With the help of MATLAB, the model was programmed and trained to recognize handwritten numbers. That data was implemented in the final circuit. Using model-based design and Simulink development environment, a custom convolutional neural network algorithm was developed for FPGA hardware implementation. The algorithm works with fixed-point data and uses stream processing for the incoming data. To implement the network, only FPGA resources are used, without using the CPU. The functionality of the neural network was verified on the DE2-115 board. The final network contains 7 layers and has a recognition accuracy of 98.6%.