Research On Highway Pavement Surface Distress Image Recognition

In recent years, as the rapid developing of expressway construction, the technology of pavement management and maintenance also needs to be improved. But the current detection of pavement surface distress is still using the traditional manual detection method. The traditional method can't meet the need of the rapid pavement condtruction, since it has many disadvantages, such as inefficiency, heavy working intensity, time-consuming evaluating process, high probability of artificial error and physical risks. Therefore, it is important to carry on the study on automatic detection and recognition of pavement distress. Based on the high-tech project of Science and Technology Department of Jilin Province–"The Development of Integrated Detection Vehicle of subgrade and pavement", this paper focuses on the technology of pavement image denosing, segmentation, extraction of information in crack image and classification of crack image based on image processing.The paper introduces the architecture of the automatic pavement distress detection and recognition system, and describe the hardware component particularly, as well as the standard and reson for selection of each hardware.Also, the principle of the system is be explained in detail, what more, the efficiency and feasibility of the scheme and algorithm used in this system is examined through simulation and test.Because of the complexity of pavement crack image (such as debris on the road, the interference around the cracks, and ununiformity of pavement because of different pavement material granularity and pavement roughness), jitter of camera in the detection vehicle and noise of electronic equipment, the difficulty of detection and recognition of pavement distress is increasing. In order to facilitate the following image analysis, we must firstly take image denoising. According to the features of crack image, the wavelet packet thresholding method for the pavement image denoising is proposed in the paper. Compared with median filter and wavelet thresholding method, this method is very effective in dealing with the image denoising. The result of image denoising by wavelet packet thresholding method shows that the method facilitates image segmentation processing greatly.The segmentation method based on fractal features is used in the paper. One of the fractal characteristics is self-similarity, which is the similarity between local structure and overall structure of system. Because the pavement crack is caused by natural factors and each image is made of single crack, the characteristic of pavement image takes on self-similarity. In the paper, we make use of fractal characteristics of pavement crack image to segment the image. After experiment, the result shows that the segmentation method based on fractal characteristics is effective and feasible.Although the segmentation method based on fractal features has a good performance for pavement crack image segmentation, the binary pavement crack image still has many noises after segmentation because of complexity of pavement crack image. And those noises are presented in the form of"isolated"points. It is necessary to eliminate these"isolated"points in order to ensure the accuracy of crack classification. A morphological method is used in eliminating"isolated"points and experimental results show that the performance of this method is satisfying.Features extracting is a key point of image recognition. Structure features of a variety of pavement distress are proposed through particular analysis of pavement distress images. The features include projection of crack and the regional statistical feature, which give a good description of the various types of pavement distress shape characteristic and show a good performance in image classification. Meanwhile the paper process the features extracting method based the fractal feature of pavement distress by calculating the fractal dimension of the pavement distress binary images. Generally, the box-counting dimension can be used to distinguish three categories of pavement distress: no crack, unidirectional crack and other crack.Finally, a classifier is designed based on RBPNN(Radial Basis Probabilistic Neural Network) and four features extracting from pavement crack image as the input and five type of pavement crack as the output. They are transverse crack, longitudinal crack, block crack, alligator crack and no crack. In order to evaluate the network, we collected 500 image samples as the training data and test data respectively. The result of experiment shows that the classifier gets a better performance to recognize unidirectional crack and no crack .The RBPNN achieves an accuracy of 95% for transverse crack, longitudinal crack and no crack,80% for block crack and alligator crack and the whole accuracy is 92%.At the same time, we also make some experiments with RBF neural network and probabilistic neural network. The respective result shows that RBPNN is better than the other two methods in recognizing pavement crack image.