Advanced bridge visual inspection using real-time machine learning in edge devices

Abstract Conventional methods for bridge inspection are labor intensive and highly subjective. This study introduces an optimized approach using real-time learning-based computer vision algorithms on edge devices to assist inspectors in localizing and quantifying concrete surface defects. To facilitate a better AI-human interaction, localization and quantification are separated in this study. Two separate learning-based computer vision models are selected for this purpose. The models are chosen from several available deep learning models based on their accuracy, inference speed, and memory size. For defect localization, Yolov5s shows the most promising results when compared to several other Convolutional Neural Network architectures, including EfficientDet-d0. For the defect quantification model, 12 different architectures were trained and compared. UNet with EfficientNet-b0 backbone was found to be the best performing model in terms of inference speed and accuracy. The performance of the selected model is tested on multiple edge-computing devices to evaluate its performance in real-time. This showed how different model quantization methods are considered for different edge computing devices. The proposed approach eliminates the subjectivity of human inspection and reduces labor time. It also guarantees human-verified results, generates more annotated data for AI training, and eliminates the need for post-processing. In summary, this paper introduces a novel and efficient visual inspection methodology that uses a learning-based computer vision algorithm optimized for real-time operation in edge devices (i.e., wearable devices, smartphones etc.).