A Study on UAV Pavement Crack Image Using R-CNN Algorithm

• Main Authors: HSIEH, HSIN-YUN, 謝欣芸
• Other Authors: WANG, LUNG-JEN
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/mrxkap

碩士 === 國立屏東大學 === 資訊工程學系碩士班 === 107 === 　　The reason of road damage is nothing more than the construction materials, construction methods and road surface often exceed the weight, and there are cracks of different sizes, some are horizontal and vertical cracks formed by road surface aging and wheel rolling; of course, there are a few that is caused by external force such as the flushing of rain. In this thesis, an effective algorithm is proposed for pavement inspection. Through the extraction of image features and the algorithm of Convolutional Neural Network (CNN), the image annotation tool (LabelImg) is used to extract features from each image, and the function of target detection is fully realized, so that maintenance personnel can identify road pits and cracks in a simple way. This work will help the pavement inspection to be carried out more effectively, which will help to detect serious damage pavements and build construction immediately. 　　This thesis uses the UAV camera to collect the pavement crack images of North Dawushan for image training. The flight height is 4 meters. The pavement crack images are divided into three categories: linear cracks, crocodile skin cracks and potholes for dangerous roads in mountainous areas. The image pre-processing includes the gray-scale image and adaptive threshold binary images. In addition, the TensorFlow environment is built and the coding is written in Python, and through TensorBoard to monitor and view the model data. Finally, the entire network model is based on R-CNN (Region-CNN), and experiments are carried out at different scales of images during the training phase, and the difference between the accuracy of the pre-processing and the non-pre-processing, etc..