Segmentação semântica de defeitos em pavimentos asfálticos com uso de redes neurais convolucionias

Abstract

The classification of defects in asphalt pavements is one of the essential factors for the correct management of the road network. Carrying out this survey manually is time-consuming, uses a considerable amount of financial resources and makes a complete inspection unsustainable. A more automatic, faster and more comprehensive alternative to this process is the application of image processing by computer vision. Recent research indicates that approaches using Convolutional Neural Networks demonstrate a good capacity to identify and classify defects in asphalt pavements. In this context, this work proposes a semi-automatic inspection method for asphalt pavements to identify and classify defects such as interconnected cracks, potholes and patches by images, employing a Convolutional Neural Network through an algorithm developed with the Python language. The research uses an Unmanned Aerial Vehicle to acquire images of asphalt pavements to compose the database carried out in two stages: images of interconnected cracks, and subsequently a complete set of images, containing the three defects. In Stage 1, 1985 images of 256 x 256 pixels were segmented, obtained at Campus I of the University of Passo Fundo, and in Stage 2, 100 images of 5472 x 3078 pixels were segmented, obtained at the same location and also on Rua Rui Barbosa, in the city of Passo Fundo/RS. The ground truth was developed using the Hasty application. The proposed Convolutional Neural Network uses the U-Net architecture associated with pre-trained networks to perform the image segmentation task. In Stage 1, the model with U-Net architecture associated with the RESNET 34 backbone presented the best performance in the segmentation of interconnected cracks. With this in mind, Stage 2 considered this model as the basis for the development of a multiclassification network with three patch variations. The best performing multiclassification network was the model that uses 32x32 pixel patches, achieving average IoU values ​​of 0.429 and average Dice of 0.526. The values ​​were considered suboptimal since the defect classes obtained inferior results in relation to the noise and pavement classes. It was also observed the need to expand the database, with balancing between classes, as a key to improving the multiclass models and enabling them to be used as support tools in defect identification.