RUB » Bibliotheksportal

Ayman Soultana, Angelina Aziz

• Object detection holds great promise as a valuable documentation tool during the operations and maintenance phase of a building, as it allows for accurate identification of equipment, potential hazards, and other critical elements within the facility. However, the process of labeling large amounts of unlabeled data can be time-consuming and impractical. Active learning algorithms offer a promising solution by selecting only the most informative images for labeling. While several studies have explored the use of active learning algorithms for object detection, many of them do not adequately address the challenges posed by class imbalances and image quality issues. This limitation becomes particularly pronounced in datasets that contain a significant number of low-quality images and a large number of classes. Thus this paper proposes an active learning algorithm that considers these issues when selecting images for labeling. The approach in this study utilizes class weighting and a method similar to diversity sampling to balance the distribution of classes in the dataset and incorporates an image quality measurement using Laplacian variance to reduce the influence of low-quality images. The approaches demonstrate sufficient results, particularly in addressing class imbalances, making it a promising solution for achieving accurate and efficient object detection.