Pavement defect detection with deep convolutional neural network

• Main Author: Chan, Zhen Yi
• Other Authors: Zheng Jianmin
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision; Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/165886
• Institution: Nanyang Technological University

Automatic inspection and defect detection using visual data such as images and videos is currently an active research topic in machine vision. The technique is widely adopted in industries, for example, for surface defect detection. The basic idea is to examine visual patterns from images of the manufactured surfaces to detect flaws. This has the advantages of overcoming the limitations of the traditional inspection approaches that depend heavily on humans and improving the efficiency and performance of the inspection. This project aims to implement a deep-learning method for early detection of potential pavement surface degradation, such as pavement cracks and faded road markings, which is important for safe driving or walking. The main purpose is to achieve fast and reliable defect detection for replacing traditional pavement defect detection. In this report, analysis and pre-processing of a large-scale dataset were conducted at first. It consists of a total of 29202 road defect images from several countries (India, Japan, Norway and Czech). Secondly, different types of state-of-the-art object detection models were implemented to achieve transfer learning for better detection precision. Lastly, different experiments have been conducted to select the best model. On top of that, different hypermeters of the best model were evaluated to achieve the best performance for the model.