Pavement Crack Image Detection And Classification

Crack detection based on digital image processing is more and more widelyapplied in the maintenance of asphalt pavement diseases. However, due to thecomplexity of the pavement crack image, it is difficult to achieve the effect of crackdetection and classification with high accuracy. To improve the shortcomings anddeficiencies in cracks extraction and classification algorithm under pavement crackdetection system, therefore, this paper carried on a thorough research.After extracting, pavement crack images are always prone to fractured andincomplete. At the same time, traditional watershed algorithm has the advantage that itcan maintain weak edge information good. Hence, this paper applies watershedalgorithm to extracting of pavement crack. In order to overcome the over-segmentationproblem appeared after crack extracting, this paper proposes the concept of gradiententropy, which is based on the characteristic that the ratio between grayscale changesteep area and total area is statistically less in pavement crack images, and serves theincrease and decrease of gradient entropy after region merging as the criterion ofjudging region merging. Afterwards, combined with that criterion, this paper designsan adaptive algorithm which can calculate the threshold of watershed edgesegmentation, and that algorithm can solve the over-segmentation problem in crackimage extracting effectively.Due to its strong ability of pattern recognition, BP neural network has beenwidely used in pavement crack classification. In the process of pavement crackclassification, there are some drawbacks, for example, network training speed is slow,error is big and fall into local minimum value easily. Aimed at these shortcomings, thispaper design and improve the composite error function to replace the traditional globalmean square error function. Meanwhile, a novel BP neural network, which can adjustlearning rate hierarchically and dynamically, is used to classify pavement crack images.The experimental results show that, compared with the traditional BP neural network, the improved algorithm has a obvious increase on the precision and speed.