STUDY OF HARDWARE-IMPLEMENTED CONVOLUTIONAL NEURAL NETWORKS OF THE U-NET CLASS

Abstract

The authors have developed and implemented two convolutional neural networks of the U-Net class: a modification of the classical U-Net and a UNet with dilated convolutions. For training and testing convolutional neural networks, data sets were used based on images from an unmanned aerial vehicle of fir trees damaged by the Ussuri polygraph. Depending on the degree of damage, the images contain trees of four classes and a background. The weights obtained during training for each of convolutional neural network were then used in the hardware implementation of the convolutional neural networks on a programmable logic integrated circuit of the Xilinx Zynq 7000 (Kintex FPGA) system-on-chip. The paper introduces the results of the study of segmentation accuracy and performance of each convolutional neural network implemented on programmable logic integrated circuits.

Keywords

• convolutional neural networks
• programmable logic integrated circuit
• hardware-implemented convolutional neural networks
• u-net