Image Segmentation With Deep Learning

Recently,artificial intelligence has been rapidly improved which benefits from large train-ing data available and computer.Deep learning has made a great breakthrough in the field of computer vision which is the most important breakthrough in the field of artificial intelligence.Deep learning is obviously better than the traditional algorithm.Image semantic segmentation is becoming increasingly important among various of computer vision tasks.It is associated with varieties of potential development in the near future such as robot intelligence,auto-driving,in-telligent medical and etc.But image segmentation is more complicated and difficult than other computer vision tasks,and its goal is to allocate each pixel a corresponding category label.Med-ical image segmentation is the most important method of intelligent medical tasks and become hotspot in computer vision tasks.However,medical image segmentation needs to design new types of deep networks due to limitations in medical image.This paper introduces the advantages and features of deep learning and recent work in deep learning,also with the differences between these deep learning networks.We explore the architecture of deep learning models and optimization methods.The significance of image segmentation for natural image processing and medical image processing is discussed.We also compare the traditional segmentation methods and the image segmentation methods based on deep learning.Details and differences of current deep networks in semantic segmentation are analyzed in this paper,including fully convolutional network(FCN),dilated convolution and deconvolutional networks.We compare these networks to those applied in classification and illustrate the advantages and which method should be still improved.In this paper,we propose a new semantic segmentation method based on ResNet and FCN with the purpose of preventing shortcomings in existing networks.Here,we combine the dilated convolution designed for semantic segmentation with residual unit to enlarge receptive field of ResNet.Meanwhile,“pre-activation”method is used in dilated residual unit.The average pool-ing pyramid and the new multi-feature fusion method are applied in this network which extracts global context information and intergrated multiple intermediate information.Fully conditional random field is applied as a "post-process" to refine boundary.Our proposed network achieves good performance on Cityscapes benchmark.We further introduce the differences and challenges of medical image segmentation com-pared to natural images.Deep networks in semantic segmentation can't be transformed in medi-cal image segmentation due to the limitation of access in medical training data and the different features of medical images.In this paper,we propose a new patch-based eonvolutional neu-ral network which makes full use of small amount data.This proposed method also convert segmentation to classification and achieve good segmentation results on brain MRI.Then we analyzed the structure and advantages of U-Net as well as the shortcomings.In this paper,we propose a new network for medical image segmentation which combines ResNet and U-Net.Our proposed network can efficiently extract features and enlarge receptive field via pyramid di-lated convolution.Our network achieve better results on ultrasound nerve segmentation dataset.We also propose a new U-Net based on Inception and a 3D U-Net,which are tested on BRATS and achieve convincing results.