چكيده به لاتين
Diagnosing and classifying surface damage in transportation infrastructure is vital to maintain security and prevent disasters. This issue is so crucial that some countries have passed approvals for periodic inspections of these infrastructures. According to these approvals, all transportation infrastructures must be regularly and periodically examined and, if necessary, be reconstructed or improved. Since performing these inspections by human agents is difficult, time-consuming, expensive, and even risky, we are looking for the automation of such inspections in this research. To perform this automation, we will take help from machine learning and machine vision methods and models, and by implementing the artificial intelligence model, we will try to automate inspections, crack detection, and classification.
In this research, we will present crack detection and classification models in two separate phases. We will offer a fast and accurate image classification model in the first phase. This model will be trained on a new combined dataset and can distinguish four distinct classes from each other. The first phase model is finally tested on a dataset that has not been seen during training, and it shows excellent results. In the second phase, we will have an extensive model that will be trained on a large dataset and used to detect cracks in images. The goal of the model in the first phase is to classify images, while the goal of the model in the second phase is to detect objects. In the second phase, to improve the performance of the model, multiple data Augmentations, different optimizations, and precise adjustment of hyperparameters have been used. Also, several images of Iranian roads have been prepared, and the model has been trained according to them to provide the best possible performance. In the second phase, we use three ways to test the model. In the first step, we will test 500 new images from Iran collected and tagged exclusively for this research, and we will achieve good results. In the second work, we prepared a video of the city and visually tested the model on it, and observed favorable results. Thirdly, we examined the model in a simulation environment to get good visual results.
We collected data and presented two new datasets to implement the two mentioned phases. In order to improve the speed of the large model in the second phase, we used optimizations so that the speed of its inference is improved and it is possible to run it in real-time on the central processor.
