Automated recognition of asphalt pavement crack using deep convolution neural network / Nor Aizam Muhamed Yusof …[et al.]
Pavement distress results in huge predicament such as environmental pollution, traffic congestion, accident and mental health. It can be classified into cracking, potholes rutting and ravelling, however cracking is the most prevalent damage on asphalt pavement. Effective and efficient pavement maint...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Universiti Teknologi MARA
2019
|
| Subjects: | |
| Online Access: | http://ir.uitm.edu.my/id/eprint/48846/1/48846.pdf http://ir.uitm.edu.my/id/eprint/48846/ https://jeesr.uitm.edu.my |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Pavement distress results in huge predicament such as environmental pollution, traffic congestion, accident and mental health. It can be classified into cracking, potholes rutting and ravelling, however cracking is the most prevalent damage on asphalt pavement. Effective and efficient pavement maintenance is crucial to identify the underlying problem, analysis of the information and selection of the most suitable rehabilitation measure. In road maintenance work, surface cracks provide insight and important information to the surveyors regarding unfavourable pavement condition in order to take effective action for maintenance and rehabilitation plan. Recently, crack identification and evaluation system using image processing technique has been proposed by several researchers to automate the manual survey process in road maintenance. However, the proposed methods often yield poor and unsatisfactory performance due the complexity of pavement texture, uneven illumination, and non-uniform background. This study proposed a deep convolution neural network (DCNN) as an alternative to image processing method to detect the existence of pavement crack in corresponding size of input image. Firstly, the study segmented the input image of the pavement into three different sizes: 28x28, 32×32 and 64×64 to produce training dataset for the network. Each training dataset is used to train the DCNN which consists of 6000 crack and non-crack patch images. Experimental results show that the highest crack detection rate was achieved by using image size of 32x32. The DCNN using this image size obtained recall, precision, accuracy and F-score of 98.7%, 99.4%, 99.2% and 99.0% respectively. |
|---|