Machine Learning Approach to Visual Bridge Inspection with Drones

This paper presents a machine learning approach coupled with image visibility optimization techniques to enable more efficient visual bridge inspection when relying solely on remote-controlled drones. This approach involved convolutional neural network (CNN) considered a representative machine learning algorithm and image analysis of optimized high-resolution imagery collected with drones. To evaluate the efficiency of this approach, remote-controlled drones were utilized to detect and measure the damage of two timber bridges located in Minnesota. Visual inspections on each of the bridges were initially performed in 2019 with two drones (DJI Phantom 4 and DJI Matrice 210), which resulted in a total of 66,572 extracted images. Then, the CNN was trained with a large number of extracted images containing varying damage types (i.e., cracking, weathering, and spalling) and attempted to classify the damage in an effective fashion. The visibility of the images that were classified per damage type through the CNN training was optimized by fine-tuning different properties of its images to take a measurement of damage specific to critical sections for each bridge. Included in the image properties were brightness, contrast, and sharpness. Through the analysis of extracted images from both timber bridges, the integrated CNN coupled with an image visibility optimization approach demonstrated the capability of improving the visibility of the damage and the accuracy of damage measurement.