| Resumo: |
Crack detection methods using neural networks have been used to analyze pavements due to their versatility, adaptability and because they have a minimal impact on road network traffic. In this context, there are still challenges to be explored, one of which is the optimization of neural network hyperparameters. Hyperparameters can be related to network architecture (filter size, network depth) or training (learning rate, dropout), and can significantly impact network performance. The present study focuses on the selection of better network architectures, and subsequent optimization of hyperparameters. 2D images of asphalt pavements were obtained using an unmanned aerial vehicle (UAV) on Campus I of the University of Passo Fundo. From images with a resolution of 5472 x 3078 pixels, cropping of 256x256 pixels, in gray scale and color, was processed and manipulated. Then, using as a basis the architecture of the MobileNet, AlexNet, DenseNet, SqueezeNet, VGG19, ResNet18, ResNet50, ResNet101 networks, the training stage of the neural networks began, which used 40,000 images, of which 80% were used to training, and 20% for the validation phase. Based on these data, a comparative study was carried out on the precision, recall, error and inference time values generated by each of the trained networks, so that it was possible to define the four best network architectures and optimize them. Finally, the AlexNet architecture was defined, with an accuracy of 97.14%, F1-Score of 86.96%, as being the best architecture choice, due to the low inference time, of around 34 milliseconds, and the effectiveness in detecting cracks, with the least use of computational power. |
|---|
