Concrete Pavement Crack Detection and Classification Using Deep Convolutional Neural Network with Grid Search Optimization

Pavement distress is the main element impacting the stability of roads and the comfort of drivers. It is crucial to identify and fix road damage promptly to prevent greater harm and decrease expenses associated with rehabilitation. Pave-ment deterioration identification has been carried out through manual means, resulting in significant time and labor requirements. Consequently, an automated method for detecting cracks is necessary to streamline this procedure. Multiple approaches exist for the automated identification of deterioration, ranging from image processing to the implementation of deep learning techniques. The process of identifying deterioration using image processing techniques often involves edge detection and threshold segmentation methods, which primarily emphasize feature extraction but remain susceptible to variations in image texture. Traditional machine learning techniques have demonstrated favorable outcomes, but they lack dependence on the features that are extracted. The application of deep learning techniques has yielded successful results in the field of distress detection, surpassing the performance of traditional methods. This research paper introduces an innovative algorithm for the identification and categorization of pavement deterioration, formulated as a multi-label classification task. In this study, images of concrete pavements were utilized as the training and test data for the models. Various types of pavement deterioration are identified and categorized, including longitudinal cracks, transversal cracks, oblique cracks, and no cracks. Moreover, in order to attain optimal performance for our algorithm, we fine-tune the hyper-parameters that compose the deep convolutional neural network model through the utilization of the grid search technique. The grid search explores every conceivable combination and selects the one that attains the greatest accuracy. Once the optimization process is finished, the effectiveness of the enhanced model is assessed using diverse evaluation metrics, including accuracy, precision, recall, and F1 score.