Multi-scale CNN Method In Image Segmentation

In recent years,deep learning has a pretty fast development,and has been widely applied in many field such as computer vision,speech recognition,text processing and so on.As one of the most popular deep network model,convolution neural network(CNN)has made a great breakthrough in capturing image features,therefore set off a wave of the research of convolution neural network's structure.In the field of image segmentation,how to capture image's feature is the key,while CNN is a pretty good choice.However,the down sampling effect of the pooling layer in CNN will reduce the resolution of the output image,leading to fail to achieve pixel level semantic segmentation.Meanwhile,lacking of enough local and global features of the target will greatly weaken the accuracy of segmentation result.Therefore,in this paper,we considering to add the multi-scale information of the image into the three dimensional dilated convolution neural network so as to increase the network's ability to capture context information of the target.Noise will seriously interfere with image segmentaion results.In order to improve segmentation accuracy,image denoising is performed before segmentation.In summary,the following three aspects are studied:(1)A image denoising method based on sparse low rank dictionary with combined sparse approximation is proposed.The similar matrix is decomposed into two part s,one is sparse part and the other is low rank part,these two parts are combined into a complete redundant dictionary.This dictionary contains priori knowledge of the image.Then the joint sparse representation method is used to restore no-noise similar matrix,it can further capture the detail information in sparse part.The experimental results show that the algorithm proposed in this paper can effectively remove the noise in the medical image and synthetic aperture radar(SAR)images,while the image details and texture information remains intact.(2)A method of venous vessel segmentation in magnetic resonance image based on migration convolution neural network is proposed.For venous vessel is small and deformable,we design a pair convolutional neural network to capture both local and global features from two scale and achieve locate and roughly segment the venous vessel.In order to increase the accuracy of segmentation result,we take the rough segment result as the initial contour of the level set model which is based on distance and contour,finally get our ultimate segmentation result.Then use the training data to train the network.The pair convolutional neural network extract target's global feature and local feature from two scales respectively.As for another case,the network's parameters of the original case can be migrated to this new case so as to saving training time.The experimental results on ten cases show that the method can effectively segment the venous vessels in magnetic resonance images and the average sensitivity is 87.60%,the average specificity is 99.96%,the average Dice similarity is 86.32% and the average modified Hausdoff distance is 0.0991 mm.(3)A multi-objective segmentation method for SAR image based on three-dimensional dilated convolution neural network is proposed.The 3D image block based on multi scale is constructed by wavelet decomposition,and the 3D image block is used as the input of the three-dimensional dilated convolution neural network.It can improve the ability of the network to capturing the local and global features of the target.The model is convenient and efficient for it adopts end-to-end structure.At the same time,the convolution kernel in 3D dilated convolutional neural network contains three-dimensional dilated convolution kernel.It can greatly increase receptive field without increasing the number of network parameters,which makes each pixel on feature map carry more information,which helps to extract global information of image.The experimental results show that this method can effectively segment ships in the SAR image,and the average sensitivity is 90.04%,the average specificity is 99.96%,the average Dice similarity is 93.86% and the average modified Hausdoff distance is 0.1110 m.