The Detection And Segmentation Of Linear Objects Based On Multi-direction Templates And Deep Learning

The detection and segmentation of linear objects is an important part of image processing.The linear objects mainly include blood capillaries in medical field,airport runways in remote sensing image and road cracks.Our highway construction has made great progress in recent years,which has greatly promoted the economic growth in related areas.One of the problems that followed was the mending and maintenance of roads.Road diseases threaten people's safety.Traditional manual detection is accompanied by subjectivity,insecurity and inefficiency.The automatic detection of road diseases has always been the goal of the researchers.The crack is the early features of most road diseases,and it has a certain length,width and direction from macroscopical view,so it's a typical linear object.In this thesis we take the crack detection as the main research content.We propose two kind of methods to realize the crack detection and segmentation base on image processing.One of them is a traditional method,we use multi-direction and large-scale templates to split the road crack from background.The other is a deep learning method,which we will use to implement the detection and semantic segmentation for road crack image.The main work is as follows:(1)We design multi-direction and large-scale templates to realize the segmentation of road crack images.Fist we proprose an adaptive threshold algorithm in pre-processing part.Then we design large scale templates with 16 directions.After that calculate the convolution between the binary image and multi-direction templates,threshold the convolution result and we will get the final result of crack segmentation.The algorithm works well both for linear cracks and block cracks.(2)We use convolutional neural network to detect the road crack.First we create our own data,the image is divided into a small pieces with size of 96?96.We put forward a new method to generate fake crack image in order to solve the problem of unbalance and insufficient samples.The training set of road image is 167,328 and tunnel image is 179,718.Then we design a convolutional neural network base on TensorFlow,the accuracy on road image is 96.01% while 74.25% on the tunnel image.Then we take mlp-convolution layer and multi-scale kernel as a reference to improve our network.The result shows that the accuracy on road image has improved to 97.02% and 89.72% on the tunnel image.Finally we visualize the feature-maps to explain why CNN can classify the crack image,and we design a single-layer convolutioanal neural network to analyze the same and differences between multi-direction templates and CNN's kernels.(3)We use fully convolutional neural network to realize the semantic segmentation of crack images.We design two networks base on PyTorch.The first network takes FCN for reference,we replace the fully connected layer in CNN with convolution layer.The network input size is 96?96.The segmentation result of FCN is relatively crude,the mean IoU is 0.5480.The second network takes U-Net for reference.The network input size is 256?256.The segmentation result of U-Net is more satisfactory and the mean IoU is 0.6514.