Investigating the optimisation of real-world and synthetic object detection training datasets through the consideration of environmental and simulation factors

• Published in: Intelligent systems with applications Vol. 14; p. 200079
• Main Authors: Newman, Callum, Petzing, Jon, Goh, Yee Mey, Justham, Laura
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.05.2022; Elsevier
• Subjects: Autonomous detection; Computer vision; Machine learning; Object detection; Optimisation; Synthetic data; Computer vision; Autonomous detection; Optimisation; Object detection; Machine learning; Synthetic data
• ISSN: 2667-3053; 2667-3053
• DOI: 10.1016/j.iswa.2022.200079

•Optimisation of a training dataset for training a deep neural network.•Identification of optimal factor levels using ANOVA and Taguchi.•Application of methodology to real-world and synthetic datasets.•Identification of optimal environmental factors for developing simulation environments. Computer vision is used for many industrial applications involving automation, especially those related to efficiency and safety. Computer vision techniques which use machine learning, such as object detectors, need a dataset of images for training and testing. Publicly available datasets or new datasets can be used. However, these datasets rarely consider whether the dataset is leading to optimal performance. Environmental factors, such as lighting and occlusion, will alter the appearance of the images and so images taken under certain condition may have different effects on training. A knowledge gap has formed as to how the test performance of deep neural networks can be improved by considering the effect and interactions of factors where either real or synthetic images are used. The following research illustrates that the different factors can have a significant impact on the test performance and demonstrates a process that can be used on real-world and synthetic images to identify the effect of each factor and discusses how this information may be used to create an optimal training dataset.