Crack length measurement is an important part of asphalt pavement detection. However, some crack measurement techniques cannot satisfy the needs of accuracy and efficiency. This study discusses application of deep convolutional neural networks (DCNN) in automatic recognition of pavement crack length in batches. Original red, green and blue images were transformed to grey-scale images to calculate their threshold and pre-extract cracks’ properties by k-means clustering analysis. Then the pre-extracted crack images were used as both training and testing samples. The process of accomplishing DCNN to recognise the crack length included the structure designing, training, and testing of the networks. The output results of well-trained DCNN were compared with those of the actual measurement to verify the accuracy of the networks. The result indicates that the training strategy including two processes overcomes the lack of crack labelled images and improves the accuracy of the network, combining with quadrature encoding and stochastic gradient descent. Recognition accuracy of DCNN is 94.36%, maximum length error is 1 cm and mean squared error is 0.2377. The error rates of length ranges 6–7 cm and 7–8 cm are bigger than other ranges Therefore, the networks can be adopted to measure the crack length accurately, but more 6–8 cm crack images should be used to improve the accuracy of the networks in future.

Many image processing techniques (IPTs) were proposed to inspect pavement defects for improving the precision and efficiency of the on-site inspections by humans. However, the various pavement conditions led to the unacceptable stability of IPTs. Therefore, an application of convolutional neural networks (CNNs) is presented for pothole detection using digital images in this study. Two CNNs, called conventional CNN and pre-pooling CNN, were designed, trained, and tested using 96,000 pavement images. Additionally, a stability study and comparative study were conducted based on the testing results. The main difference between the two CNNs was that a pre-pooling layer was used in the pre-pooling CNN to pre-process pavement images before the first convolutional layer. The results demonstrated that the optimised pre-pooling CNN had the 98.95% recognition precision in the testing. The stability study indicated that the optimised CNN model had the robustness in various real-world situations (e.g. light conditions and pavement materials). Compared with the traditional IPT methods, the CNN had a higher precision for extracting pothole features autonomously.