Improved Image-Based Super Resolution and Concrete Crack Prediction using Pre- Trained Deep Learning Models

Abstract

Detection and prediction of cracks play a vital role in the maintenance of concrete structures. However, manual inspection is not advisable in case of skyscrapers, underground structures, bridges and dams. These conditions result in having images captured from different sources wherein the acquisition of such images into the network may cause an error. The errors are rectified by a method to increase the resolution of those images and are imposed through Super-Resolution Generative Adversarial Network (SRGAN) with a pre-trained model of VGG19. After increasing the resolution then comes the prediction of crack from high resolution images through Convolutional Neural Network(CNN) with a pre-trained model of ResNet50 that trains a dataset of 40,000 images which consists of both crack and non-crack images. This work makes a comparative analysis of predicting the crack after and before the super-resolution method and their performance measure is compared. Compared with other methods on super-resolution and prediction, the proposed method appears to be more stable, faster and highly effective. For the dataset used in this work, the model yields an accuracy of 98.2%, proving the potential of using deep learning for concrete crack detection.