A Machine Vision Based Method For Concrete Crack Detection

With the development of economy and society,a large number of concrete facilities have been put into construction and operation,and the demand for defect detection of concrete facilities is increasing day by day.Among them,cracks are the most common defect in concrete construction.Cracks will make it easier for corrosive substances such as carbon dioxide to infiltrate into the interior of the concrete,causing hidden dangers in the durability and safety of the concrete facilities,and even causing serious structural damage.However,the current detection methods are still based on manual detection,which is costly and subjective.And the detection efficiency is easily affected by the degree of fatigue of the inspectors,which causes the frequent occurrence of missed detection.Moreover,manual inspection is very dangerous in many situations.With the rapid improvement of computer hardware performance and the emergence of drones and other automatic cruising equipment,along with the fast development of computer vision technology,the automated detection of concrete cracks has become possible and this field has become a hot research field.This paper first analyzes the status and trends of crack detection technology at home and abroad.Then this paper summarizes the characteristics of the current research works,and proposes a novel and uniform concrete crack recognition framework based on computer vision technology.This method is not only very robust and has high detection accuracy,but also will also exert an important and positive significance for the development of automated and intelligent concrete quality detection technology.(1)First of all,the collected images are preprocessed by graying.Gabor filters are then used for filtering and extract multi-directional and multi-scale features.(2)The extracted Gabor features are encoded using uniform LBP(Local Binary Pattern),and ULGBPHS features(Local Gabor binary pattern histogram sequence)are extracted as image descriptors.(3)Since dimension of the extracted ULGBPHS features is large and the features include a large number of nonlinear modes,the Kernel PCA method is used in this paper for nonlinear feature compression in order to improve the performance of the classifier.(4)Accurate classifier is needed for identification purpose.This paper proposes using an Oblique Random Forest classifier based on Proximal Support Vector Machines.(5)Because the parameters of the oblique random forest classifier were previously set according to the experiences of the researcher,which is subjective and inaccurate.This paper proposes using adaptive genetic algorithm to optimize the parameters.In the experimental evaluation section,this paper discusses the scale and direction parameters of Gabor filter and the block size parameter of LBP on the recognition rate.At the same time,we use the t-distributed stochastic neighbor embedding algorithm to visually study the data structure characteristics of Kernel PCA and PCA compression methods and discuss the impact of different compression methods on the recognition rate.At the end of the experiment section,the classifiers used in this paper were compared with typical traditional classification methods.The results verified the validity and accuracy of the Proximal Support Vector Machine based Oblique Random Forest classifier for the identification of concrete cracks.The concrete crack detection method designed in this paper will exert positive impact for promoting the development of intelligent concrete inspection technology.