Dataset for "Single Board Computing for Image Recognition on an Autonomous Drone"

This dataset provides the testing data utilised within the conference paper "Single Board Computing for Image Recognition on an Autonomous Drone" published by IEEE.

Article Title: Single Board Computing for Image Recognition on an Autonomous Drone
Article authors: Siu Chung Kevin Shek, Joseph Butterfield, Adrian Murphy & Ivor Spence
Article DOI: 10.1109/IEACon57683.2023.10370368

Abstract for dataset:
The dataset consists of 3 different folders related to the experiments performed within the paper mentioned above. The folders contain set of images of square targets that contains a smaller square target that contains within it with an alphanumeric character.

Abstract for paper:
Image recognition is a well-established method of extracting information from digital images. However., the selection, testing and integration of such methods when designing the software and hardware to enable situational awareness for autonomous systems in unstructured environments, remains a challenge. This is particularly the case when lightweight and low- cost are the driving requirements for the system. The main aim of this paper is to compare the k-nearest neighbour (K-NN) traditional machine learning method against a deep learning model for a character recognition task intended for implementation on a single-board computer on an aerial drone. A series of image recognition experiments were carried out which attempted to identify a set of alphanumeric characters using a Raspberry Pi (RPi) image capture system (RPi V2.I camera with an RPi 4 B) with a total cost of £97.50 and weight of 49g. The experiments identified that (LSTM) within Tesseract optical character recognition (OCR) was 2.I2% more successful on average than a K-NN algorithm for character recognition but took 58.30% longer on average to process a single image. It also indicated that for human-defined shape detection in an unstructured environment, data specific to each search environment must be used. This highlighted the importance of tailoring hardware and software selection to the application.

Information about the dataset can be found within the zip file. It contains 3 types of testing dataset, a lux reading excel spreadsheet and a README file. Please download all the zip folders first, next open the "square_target_testing_dataset.zip" to be able to access the corresponding 2 dataset parts (square_target_testing_dataset.z01 & square_target_testing_dataset.z02) to be open with Explorer or double-click.

本数据集为发表于IEEE的会议论文《Single Board Computing for Image Recognition on an Autonomous Drone》（自主无人机图像识别的单板计算）所使用的测试数据。

论文标题：Single Board Computing for Image Recognition on an Autonomous Drone（自主无人机图像识别的单板计算）
论文作者：Siu Chung Kevin Shek、Joseph Butterfield、Adrian Murphy、Ivor Spence
论文DOI：10.1109/IEACon57683.2023.10370368

数据集摘要：
本数据集包含3个与上述论文中所述实验相关的文件夹，文件夹内存储带有方形靶标的图像集，该方形靶标内部嵌套有一个更小的方形区域，区域内包含一个字母数字字符。

论文摘要：
图像识别是从数字图像中提取信息的成熟方法。然而，在非结构化环境中为自主系统设计具备态势感知能力的软硬件时，此类方法的选型、测试与集成仍是一项挑战。当系统以轻量化与低成本为核心设计要求时，这一问题尤为突出。本论文的核心目标是，针对拟在无人机单板计算机上部署的字符识别任务，对比传统机器学习方法k近邻（k-nearest neighbour，K-NN）与深度学习模型的性能。本研究开展了一系列图像识别实验，旨在通过总造价97.50英镑、总重量49克的树莓派（Raspberry Pi，RPi）图像采集系统（搭载RPi V2.I摄像头与RPi 4 B主板）识别一组字母数字字符。实验结果表明，在Tesseract光学字符识别（Optical Character Recognition，OCR）框架内的长短期记忆网络（Long Short-Term Memory，LSTM）在字符识别任务中的平均准确率较K-NN算法高出2.12%，但单张图像的平均处理时长较K-NN多出58.30%。实验同时表明，在非结构化环境中开展人工定义的形状检测时，必须使用针对对应搜索场景的专属数据集。这一结果凸显了针对具体应用定制软硬件选型方案的重要性。

数据集的详细说明可在压缩包中获取，该压缩包包含3类测试数据集、一份光照读数Excel表格与一份README文件。请先下载所有压缩包文件，随后双击或通过资源管理器打开「square_target_testing_dataset.zip」，即可访问对应的两个数据集分片文件（square_target_testing_dataset.z01与square_target_testing_dataset.z02）。