Research On Image Recognition Algorithm For Complex Crack Diseases In Subway Tunnels

At the same time of the rapid development of urban rail transit in China,the safety inspection of subway tunnel infrastructure require automated and intelligent theory and technology research urgently.The subway tunnel lining environment is harsh,there are a lot of noise and uneven illumination,and their texture features are very complex.The traditional manual detection and image recognition methods have been gradually replaced by deep learning algorithms that are more intelligent and effective.Therefore,a subway tunnel lining image recognition algorithm was proposed in the thesis combined with image processing and deep convolution network,and sample libraries of subway tunnel disease images were established.It can distinguish crack images from non-crack images quickly and accurately,and can achieve target intelligent search and category labeling in complex images.A theoretical model of pixel-level shallow image processing was constructed in the thesis including image preprocessing algorithm,image multi-level feature analysis algorithm based on connected region and circumscribed rectangle extraction of feature texture connected region,which can filter out noise and discriminate crack region.4 preliminarily classified image sample libraries were established.The classification algorithm combined with image processing and deep convolution neural network extracted and analyzed different regions of the sample libraries to realize the classification of crack images and non-crack images,using the improved Alexnet deep convolution network.And the traditional classification algorithm such as SVM was compared and analyzed in detail.In order to realize disease intelligence detection more precisely,a target detection algorithm based on deep convolution network without any pixel operation was designed in the thesis,and a target annotation sample library about subway tunnel lining images was built Intelligent identification and category marking of targets such as crack were accomplished using an improved SSD full convolution network structure.The training accuracy of classification algorithm based on deep learning about crack images and non-crack images is up to 98.6%,and the test accuracy is as high as 97.8%,which is better than traditional classification algorithm.And the sample library combined with pixel-level shallow image processing and circumscribed rectangle extraction of connected region is the best.The mAP value of target detection algorithm based on deep learning about complex images is 0.531,the AP values of crack and leakage are 0.424 and 0.718 respectively,and the detection rate are as high as 93.08%and 94.10%.It has certain application value in the actual and intelligent identification of subway tunnel diseases,and provides technical support and research ideas for the application of deep learning in the subway tunnel diseases detection.