MobileCrack: Object Classification in Asphalt Pavements Using an Adaptive Lightweight Deep Learning

Abstract

One of the key steps in pavement maintenance is the fast and accurate identification of the distresses, defects, and pavement markings and ability to conduct the maintenance before the irreversible damages. Recently, convolution neural network (CNN) has emerged as a powerful tool to automatically identify the pavement cracks, where many of the CNN models take long computation time. To solve the problem, an adaptive lightweight model, named MobileCrack, is proposed in this study. MobileCrack realizes the fast computation using the following settings: (1) reduce input image size, where besides the original input images, images with a side length of 1/2,1/4,1/8… of that of the original square images are also input into the model by using the resize command; (2) group convolution is used; and (3) global average pooling is used because it normally has less parameters compared with the fully connected layer. MobileCrack will then compute the combinations of different resized input images and different neural network structures, to find the optimal reduced image size and neural network structure with satisfactory accuracy using reasonable computation time. To verify the applicability of MobileCrack, 10,000 input images with size 400×400 are trained for the classification task of crack, sealed crack, pavement marking, and pavement matrix. Based on the computation results of combinations of images with different sizes (400×400, 200×200, 100×100, and 50×50) and different stacking numbers of core modules n (3, 4, 5, and 6), the optimal model is determined as image size 200×200 and n=4, where the test accuracy is 0.865 within reasonable computation time (runtime=47 ms to test one image). This optimal model will be automatically used for further tests with image size 400×400 for a fast computation, which realizes the lightweight adaptive goal. Results also show that the test accuracy of MobileCrack is higher than that of the AlexNet and visual geometry group (VGG), and the parameters of MobileCrack are approximately 1/4 of that of the classic lightweight CNN model MobileNet, which saves the storage space. It is concluded that that the proposed adaptive lightweight CNN model, MobileCrack, can be used for the fast object classification on asphalt pavement crack images.