Road distress identification using deep machine learning algorithm

Abstract

This study focuses on developing an automated road distress identification and classification framework using state-of-art edge detection and deep machine learning technique. In this study, the Convolutional Neural Network (CNN) and Sobel Edge Detection (SED) have been harmonized to automatically identify and classify road distresses using a moving camera. The Sobel Edge Detection techniques have been used to determine the 2-D spatial gradient of a pavement image to identify the distress from it. The pavement image has been converted into machine readable binary image, where the distress area contains a specific identification flag. This identification flag has then been recognized and categorized by blob analysis. A bounding box with measurable dimensions has also been created using the Blob analysis. The image within the bounding box has been given as input in the CNN architecture for classification. A layer of CNNs consists of three main sublayers, which include: convolutional layers, pooling layers, and fully connected layers. In this research, two types of data have been collected: (i) Static image; and (ii) Dynamic image. Three thousand five hundred static pavement images having a resolution of 3264 ×2448 were collected from different streets and highways within the Dhaka region. The dynamic image dataset is divided into two parts: (a) Synthetic data; and (b) Real-time data. An artificial video using Macromedia MX has been made and used as synthetic data. At the same time, two real-time videos were collected from DIT Road and Dhaka Mymensingh Highway from coordinates 23.7547° N, 90.4154° E to 24.3654° N, 91.1641° E. The static dataset has been enriched using the image augmentation technique. Training of the CNN model is done using the randomly selected static image data. The hyperparameters (i.e., number of layers, number of filters, number of epochs, initial learning rate, and percent of training data) have been tuned using the graphical optimization technique. The synthetic dataset has been used to estimate the initial guess value of the hyperparameters. The optimum parameter value has been found to be: Number of layers = 2, Max Epoch = 9; Training Data = 70%; Learning Rate = 3.64E-05; Number of Filters = 30. The training and testing accuracy were found to be 99.22% and 98.78%, respectively, with the optimum hyperparameters. The hyperparameter optimization process involved a total of 992 hours of processing time. Confusion Matrix (CM) and Receiver Operating Characteristics (ROC) analysis were done over the trained and tested results and the analysis shows consistent efficacy in classifying each distress accurately. The results have been compared with the baseline method Support Vector Machine (SVM). The comparison shows that SVM achieved 94.2% and 80.3% accuracy in distress identification and classification, respectively. In contrast, the CNN model has achieved 98.1% and 97.7% accuracy in distress identification and classification, respectively, which shows that the developed CNN model performs better than the baseline method SVM. Pavement Relative Scoring has been conductive which is indicative of pavement health. Finally, A tool is developed named Road Distress Training and Classification (RoadDisTrac) using MATLAB runtime environment.