Pavement Crack Extraction Algorithm Based On Deep CNN

With the development of China’s transportation system,the normal operation of highway systems and the protection of traffic safety have become increasingly important targets for routine maintenance of transportation system.Pavement crack detection is the key content of pavement distress detection.The method of crack detection has experienced the process from manual detection to automatic detection.In recent years,convolutional neural networks have dominated the field of computer vision.With its complex network structure and learning capabilities,and with the rapid development of computing hardware and accumulated large data,it has brought unprecedented progress to image recognition algorithms.In order to solve the problem of automatic detection of cracks and other diseases in the two-dimensional gray-level image of pavement,a set of preprocessing,region determination,semantic segmentation and multi-layer feature integration based on deep convolutional neural network is proposed for the pavement images collected by line scan camera.First,a new method for solving inhomogeneous illumination was proposed for pavement image acquisition platform based on line scan camera.The method analyzed the distribution of uneven illumination and extracted a precise correction mask by processing every pixel column respectively.After correcting with the mask and stretching,the result exhibited sound processing effect.Compared with the available correction algorithms,the proposed method eliminates the abrupt changes of grey value and retains most of the information after image processing,and is proved to be suitable for pavement distress detection system with line scan camera.Then data enhancement and normalization is applied to prepare data for neural networks.Second,a two-step CNNs method is proposed to detect crack-pixels from pavement pictures and to reduce the time consumption.The method contains two main parts: CNN-1 for patch classification and CNN-2 for semantic segmentation.The first part chooses regions with high probability to contain cracks and send them to CNN-2 to get pixel-wise detection results.The results show that the obtained precision,recall and F-1 score are 0.78,0.73,0.75,and compared with pure-segmentation network CrackNet,the two-step CNNs method reduce the processing-time on one test image from 209 ms to 1.80 ms while the loss of accuracy is small.Last,this paper studies the multi-layer feature integration method based on transposed convolution network.In the two-step convolution neural network proposed in Chapter 4,the first step classification network is only used for the classification results of the last layer,in fact,the inter layers contain a lot of feature information,especially the deeper layers’ receptive fields are larger,and the global abstract features contained in it can be used as a supplement to the extracted information of the segmentation network.A set of transposition convolution networks are train which integrate the inter layers of classification network and the output layer of segmentation network.Through comparison and selection,it is find that the transposition convolution networks which integrate multiple high-level outputs show the best performance.It is proved that the integrated results can extract crack information more accurately,especially retain the continuity characteristics of cracks.