A group of researchers from the University of Electronic Science & Technology of China and Lutgert College, Florida Gulf Coast University, USA are working on a project that will help in
Civil Infrastructure Health Assessment
Their project proposes a multi-task framework, EnsembleDetNet which leverages object detectors and classifiers based on ensemble learning for damage detection and scene classification
They introduced a novel attention module that brings significant improvement in EnsembleDetNet. Their experiments demonstrate the practicability and effectiveness of EsenmbleDetNet.Their comparisons with state-of-the-art detectors and classifiers indicate EsembleDetNet is favorable under variant evaluation metrics
Below is the abstract of their research;
Damage detection pre and post-construction of civil infrastructures aids to ascertain the health status of infrastructures. Recently, deep learning and smart devices such as drones are widely being utilized in the detection of damages on infrastructures towards autonomous structural health assessment. However, challenges including object scale variation due to drones’ movement and sparse scenes from the percept of drones make autonomous structural health assessment challenging. Furthermore, assessing the health status of infrastructures should not be limited to the local pictorial view but rather a global view to estimate the extent of damage to the infrastructure. Given that, this paper proposes a multi-task framework, EnsembleDetNet which leverages object detectors and classifiers based on ensemble learning for damage detection and scene classification. EnsembleDetNet induces diversity, and strength-correlation to enhance robustness in detecting damages and the level of damage in local and global pictorial view respectively. Further, a novel attention module proposed brings significant improvement in EnsembleDetNet. Extensive experiments with a public dataset and an onsite verification utilizing a micro drone demonstrate the practicability and effectiveness of EsenmbleDetNet. Lastly, comparisons with state-of-the-art detectors and classifiers in both detection and multi-label scene classification tasks indicate EsembleDetNet is favorable under variant evaluation metrics.
Keywords: Deep learning, structural health monitoring, detection, autonomous, ensemble, unmanned aerial vehicles, drones.
