Enhancing road crack detection with Neural Architecture Seeks Large Neural Network: Leveraging deep learning and Augmented Minority Over-Sampling Technique on public and custom developed datasets

Abstract

Deep neural networks for identifying road cracks have emerged as a crucial field of study, marking a significant advancement in infrastructural maintenance. The proposed research presents a novel Neural Architecture Seeks Large Neural Network for detecting road cracks, featuring 27 convolutional layers and ten modules, leveraging Softmax for classification. Initially, the custom developed dataset contained 30,283 images, expanded to 218,073 images using the Augmented Minority Over-Sampling Technique. For better comparison, only 30,350 images are utilized from this expanded data set. Similarly, the Karlsruhe Institute of Technology and Toyota Technological Institute dataset grew from 30,274 to 217,972 images after Augmented Minority Over-Sampling Technique processing. However, 30,274 images from the original dataset and 30,327 images from the Augmented Minority Over-Sampling Technique dataset have been processed. This normalization process aimed to ensure a balanced comparative study between the original and augmented datasets, minimizing differences and enhancing the reliability of results across the datasets. Utilizing a 70/30 train-test split, the network effectively classifies seven types of crack anomalies. The model achieves 83.7% and 89.8% accuracy on the original and augmented Karlsruhe Institute of Technology and Toyota Technological Institute datasets. The custom dataset reaches up to 91.0% accuracy for post-augmentation, while the pre-augmentation accuracy is 90.7%.