Recognition of underlying surface using a convolutional neural network on a single-board computer

Abstract

The results of the development of hardware and software system (micromodule), which detects and classifies underlying surface images of the Earth are presented. The micromodule can be installed on board of a light unmanned aerial vehicle (drone). The device has the size 5.2×7.4×3.1 cm, the weight52 g, runs on a Raspberry Pi Zero Wireless single-board microcomputer and uses a convolutional neural network based on MobileNetV2 architecture for real-time image classification. When developing the micromodule, the authors aimed to achieve a real-time image classification on inexpensive mobile equipment with low computing power so that the classification quality is  comparable  to  popular  deep  convolutional  network  architectures. The provided information could be useful for engineers and researchers who are developing compact budget mobile systems for processing, analyzing and recognition of images.