Crack detection and classification in asphalt pavement images using deep convolution neural network

• Published in: Proceedings - 8th IEEE International Conference on Control System, Computing and Engineering, ICCSCE 2018
• Main Author: Yusof N.A.M.; Osman M.K.; Noor M.H.M.; Ibrahim A.; Tahir N.M.; Yusof N.M.
• Format: Conference paper
• Language: English
• Published: Institute of Electrical and Electronics Engineers Inc. 2019
• Online Access: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85065018963&doi=10.1109%2fICCSCE.2018.8685007&partnerID=40&md5=3ee52440bf031a9ff0b5611a296ff2ba

Pavement distress particularly cracks, are the most significant type of pavement distress that has been studied for many years due to the complicated pavement crack condition. The continuous severity of crack can cause a dangerous environment that may affect the road users. Therefore, an efficient computer algorithm plays an important role in developing analysis tools for automated crack detection. In Malaysia, many of road surveyors are still employing manual inspection, which is a labour-intensive, error-prone and hazardous task. Various attempts have been made to automate this task by using image processing techniques. However the method turns out to suffer from the problem of lighting variation and complexity of the background such as low contrast on the surrounding pavement that similar to the intensity of crack. This study proposed a deep convolution neural network (CNN) as a detection system of ashpalt pavement crack that capable to detect and classify the pavement crack robustly when dealing with complexity background image. A digital camera is used to capture the image of pavement crack. Then, the captured images are divided into two (2) different grid scales, 32× 32 and 64× 64, and further fed as input to the first deep CNN. For each grid size, the network is trained independently to detect the presence of crack in the image. In the classification stage, the captured images are binarized with the similar grid scales to extract the crack pattern. The binary images containing two types of crack, transverse and longitudinal are then fed as input to the second deep CNN and trained to identify the type of crack. Experimental results show that deep CNN using 32x32 grid scale images provides higher performance for crack detection and classification compared to 64x64. The network achieved the recall, precision and accuracy of 98.0%, 99.4% and 99.2% respectively for crack and non-crack detection, while the performance for transverse and longitudinal achieved the accuracy of 98% and 97% © 2018 IEEE.