Pavement Crack Detection Using Otsu Thresholding for Image Segmentation Amila Akagic, Emir Buza, Samir Omanovic, Almir Karabegovic University of Sarajevo, Faculty of Electrical Engineering, Department for Computer Science and Informatics Zmaja od Bosne bb, Kampus Univerziteta, 71000 Sarajevo Email: {aakagic, ebuza, somanovic, akarabegovic}@etf.unsa.ba Abstract—Pavement cracks are the first signs of structural damage in the asphalt pavement surfaces. The oldest method for detection and estimation of the pavement cracks is human visual inspection, also known as manual visual inspection. However, using human inspectors is very time consuming, very expensive and poses a risk to human safety. Another negative side is the fact that the task generally requires road to be closed. Hence, automatic prevention and reparation of cracks on the asphalt surface pavements is an important task, especially because the advanced stages of road deformation lead to formation of potholes. This has negative impact on the total reparation cost. In this paper, we proposed a new unsupervised method for the detection of cracks with gray color based histogram and Ostu’s thresholding method on 2D pavement image. At first, the method divides the input image into a four independent equally sized sub-images. Then, the search for cracks is based on the ratio between Ostu’s threshold and the maximum histogram value for every sub-image. Finally, all sub-images are assembled into the resulting image. The method was tested on the dataset which contains different pavement images with very versatile types of cracks. The results showed that the proposed method achieves satisfactory performance, especially in the cases of low signal-to- noise ratio, and is very fast. Keywords: Cracks detection, Unsupervised method, Image processing, Image segmentation, Computer Vision I. I NTRODUCTION Image-based pavement health monitoring techniques have a history of more than 30 years [1]. Techniques range from early windshield surveying to manual and/or computer aided analysis of pavement surfaces digital images. Currently, it is common practice to capture and analyze 2D digital images with the high-speed cameras mounted on a specialized or passenger vehicles. The analysis is usually performed off-line on an as-needed basis, or online to obtain current information. Today, many pavement maintenance agencies employ the semi-automated and the fully-automated image-based meth- ods for collection of the pavement state data. For example, the national roads in France are periodically inspected in a three years cycle [2], while in Germany the relevant surface characteristics are inspected in a four year cycle [3]. The analysis of pavement images should reveal various existing distresses on asphalt pavement surfaces. They are usually classified as high- or low-level severity distresses. Former are commonly known as potholes [4]–[6] and cracks, while latter are patches [7], surface deformations and defects. Cracks on asphalt pavement surfaces are the most common distress type. They represent important indication of possible structural damage that may lead to dangerous situations. Many algorithms with different approaches exist in literature [1], [8]–[11], however the advancements in development of accurate, automated, low-cost, effective and reliable image- based pavement cracks detection are still needed. 2D image analysis of any type of distress is a challenging task due to limiting range of information that can be extracted from 2D space. Cracks on images appear as thin, irregular, darker intensity lines, surrounded by strong textured noise. Technologies such as 3D imaging and texture analysis can provide additional information, but they are considered ex- pensive in two ways: use of expensive equipment and very high demand for computational processing. Many supervised algorithms exist in literature [12]–[17], however very few fully unsupervised [18], [19] methods exist. In this paper, we address the problem of efficient automated pavement cracks detection. We use unsupervised vision-based approach to analyze photometric data from 2D asphalt surface pavements images and extract important information to per- form efficient image segmentation. The method consists of two major phases. In the first phase, the original image is sliced into a four sub-images, then arithmetic mean and standard deviation of each sub-image is calculated. In the second phase, we calculate histogram and Otsu’s thresholding for each sub- image. Then, we define the smallest ratio between minimal and maximal absolute difference between these two values for all four sub-images. This information is then used to eliminate the background pixels and effectively select pixels that represent cracks. We have performed the comprehensive tests on the real pavement images with cracks at different severity levels. Tests show promising results with very low signal-to-noise ratio and very fast execution time. The method is applied without pre- or post-filtering and without any knowledge about shapes of cracks. The method is suitable for parallel execution since it can take advantage of spatial distribution of crack pixels and it can analyze four sub-images in parallel. This paper is organized as follows: the related work is briefly reviewed in Section II. In Section III, a new method for detection of high severity distresses on the asphalt pavements has been proposed. The implementation details, data set and experimental results have been presented in Section IV. The paper is concluded with some remarks in Section V. 1262 MIPRO 2018/CIS