Convolutional Neural Network and Interpretable Deep Learning for Concrete Crack Image Classification

Abstract

The development of unmanned aerial vehicles (UAVs) makes automation of visual tasks possible, such as crack detection. Crack detection has many challenges and, in this work, the utilization of image processing through deep learning-based computer vision is conducted for concrete surface image classification. The widely used deep learning architecture for computer vision is the convolutional neural network (CNN). This paper discusses the creation of CNN models for concrete crack image classification and the role of interpretable deep learning in the model’s evaluation. Three convolutional architectures combined with a proposed classifier architecture were trained and evaluated quantitatively and qualitatively. The quantitative evaluation metrics are precision, recall, F1-score, and accuracy. The qualitative evaluation concerns the feature highlight of the model using SmoothGrad. The result is that even though the model with InceptionV3 has the best quantitative classification metric values (accuracy of 96%), the one with EfficientNetV2S has the best feature highlight. Thus, the model considered the best is the one with EfficientNetV2S since the accuracy is already considered high (94%). This highlights the importance of qualitative evaluation on a deep learning-based computer vision model to ensure the correct feature is considered as the deciding point of classification.