Experimental Evaluation of Convolutional Neural Networks in Asphalt Concrete Computed Tomography Scan Image Analysis

This research proposed a Convolutional Neural Network (CNN) model for Asphalt Concrete Computed Tomography (CT) Scan image classification, which uses SoftMax activation for the output layer to yield the Air Void prediction with a testing accuracy of 0.9684, weighted average precision, recall, and f1-score are 0.97 in the early stopped model, and 0.99 in the best model (smallest validation loss). Moreover, by switching to Linear activation, the CNN model can generate numerical values for Air Void, which has the average percentage difference of 0.42%, −1.47%, 2.87%, 2.3%, and 2.27% for Air Void contents of 10.4%, 15.6%, 18.9%, 21.9%, and 25.6%, respectively. Furthermore, by integrating the CT image with Grad-CAM++ generated heatmap, instance segmentation of the Asphalt Concrete CT image as components of Air Void, Binder, and Aggregates is more effective than using image only. The application of using segmentation results for Asphalt Concrete 3D modeling is also presented.