Deep Learning Methods for Classification of Road Defects

Abstract

Many methods are proposed in the classification of pavement defects through the extraction of features from data and the use of machine learning algorithms to solve the problem. But there are still limitations such as training time, accuracy and the sensitivity of the system to environmental conditions. This research proposes the method of optimizing the automatic classification system of pavement defects based on Convolutional Neural Network, a deep learning method commonly used and highly effective in artificial intelligence and digital image processing. This system is guaranteed to operate stably in limited light conditions, shading and complex-shaped defects. Our experiment is performed with 3 data sets (INESC TEC - Portugal, Irkutsk - Russia, Thai Nguyen - Viet Nam). The data obtained from VGG-16 method is compared with data obtained from the Random Forest algorithm and Support Vector Machine method. The experimental data show that the proposed method allows the Random Forest algorithm to work faster, more stable and gives more accurate results. The result after classification is 97.07 %, 97.62 %, 98.50 %) respectively.