Site-optimized training image database development using web-crawled and synthetic images

Since most state-of-the-art vision technologies have recently originated from machine learning or deep learning algorithms, it has become very important to build a large, high-quality training database (DB). To this end, this paper proposes an automated framework that creates images using web crawling and virtual reality techniques, labels target objects, and generates a training DB for vision-based detection models. The framework contains three main processes: (1) image collection and labeling using web crawling; (2) image producing using a 3D modeling tool; and (3) foreground-background cross-oversampling. As a result, the framework constructed a training DB composed of 99,800 images in 42 min. The deep learning model was trained by the generated DB and showed macro F1-scores of up to 96.99%. These results imply that the framework successfully constructed a high-quality training DB within a short period of time. The findings can contribute to reducing time and effort in developing vision-based monitoring technologies.