Advanced crack detection and quantification strategy based on CLAHE enhanced DeepLabv3+

Abstract

Deep learning method has been developed for pavement crack detection. The improvement in detection accuracy is limited by the image quantity and detection algorithm in engineering practice. This study proposes an innovative method to efficiently identify pavement cracks by moderating the training process. The dataset was established after a sensitivity analysis by considering the shooting height and image volume. The DeepLabv3+ model was used to train the originally collected images, and the CLAHE-augmented images were used for prediction. The result showed that the detection accuracy was enhanced by an average of 1.51%. This new method was validated in two scenarios, leading to average accuracies of 95.07% and 96.61% in detecting linear and alligator cracks, respectively. This detection process avoids repeatedly investing new images in model training, accelerating crack recognition by augmenting the images to be detected. The proposed method fulfills the requirements of engineering practice, based on which a road maintenance strategy is proposed.