Adaptive Thresholds Pavement Crack Detection Method Based On FRIT And Random Structured Forest

With the improvement of China’s living standards,the scale of public infrastructure also gets an expansion.As an important part of public infrastructure,roads provide irreplaceable convenience for citizens’ travel and transportation.However,as the increase of highway mileage year by year,the maintenance task needs to be solved urgently.Since the crack is a direct manifestation of pavement decay,identifying cracks in time is an important part of pavement maintenance task.In the early years,crack detection mainly relied on manual detection.However,because of the increase of highway maintenance mileage,it is impossible to achieve real-time and effective crack detection only relying on manual method.Therefore,timely and effective detection of pavement cracks has become an important task of pavement maintenance.In nowadays,the rapid development of image processing and computer technology make utilizing image processing for pavement crack detection possible,and using image processing to detect pavement crack also gets an unprecedented development.However,noises around cracks such as oil and water stains,the complicated shape of cracks and the heterogeneous strength of the cracks may cause the traditional crack detection algorithm cannot achieve high accuracy recognition of pavement cracks.Referring to the existing crack detection algorithms,this thesis proposes an adaptive double-thresholds pavement crack detection method based on the finite ridgelet transform and random structured forest to realize the accurate identification of pavement cracks,and carries out experimental verification and analysis.The main work of the thesis includes:1)Aiming at the problem of insufficient contrast between cracks and background caused by the heterogeneous strength and complex noisy background including oil,a crack enhancement algorithm based on the finite ridgelet transform is proposed.The contrast between cracks and background is enhanced by multiple enhancement of whole coefficients who are of the co lumn of minimum coefficient in the finite radon domain.The noises are smoothed by multiple suppression operation of detail coefficients of each layer of wavelet transform in ridgelet domain.After the enhancement algorithm,more suitable crack images are produced for feature learning and training classifier design.2)Aiming at the characteristic of complex crack topology,this thesis proposes a method to generate crack score map by using random structured forest.The feature representation of crack block is studied in detail.Taking advantage of channel features and pairwise difference features,features of crack block are extracted from different aspects such as color,gradient amplitude,direction and pairwise difference.The knowledge of random structured forest is studied,including samples’ selection,structured learning,labels’ discretization and so on.Fixed size crack blocks selected on the crack images and the relevantly extracted features are putted into the random structure forest for learning to generate a crack detector which can predict various crack shapes and generate crack score map.3)According to the characteristics of crack score map obtained by random structured forest,an adaptive double-thresholds crack extraction algorithm is proposed.By setting a low threshold to exclude the background pixels with low score,and setting a high threshold to extract the crack pixels with high score.Besides,an iteratively traversing method of the remaining indeterminate pixels’eight neighborhoods is proposed to extract the crack pixels in the remaining indeterminate pixels from the complex background noise pixels.And a cascaded morphological method is used to enhance the connectivity of the cracks.Finally,the experimental results show that the proposed method can get better detection results than other algorithms in the case of complex crack background and topology.