Research Of Image Segmentation By Deep Learning

Image segmentation is one of the main research directions in the field of digital image processing.It plays a very important role in many computer vision applications.Because of the importance and difficulty of image segmentation,image segmentation technology has been paid more and more attention by researchers since 1970 s.Although many methods have been proposed to solve these problems,there is still not a segmentation algorithm which is suitable for most images.Most of the traditional image segmentation methods are first extracting features from the image,and then mapping them to a model.Not only the process is complex,but also the results are not robust,and can not often give semantic information.Therefore,it is of great significance to improve the efficiency of image segmentation algorithm and improve the segmentation effect.Thanks to the development of big data and high-performance computing devices,the deep learning has become a research hotspot in the field of artificial intelligence.Deep learning is essentially the development of neural networks,and has made a breakthrough in speech recognition,face recognition,object recognition,target detection and many other fields of artificial intelligence.One of the most widely used models of deep learning is convolutional neural networks in the field of computer vision.Convolutional neural networks can reduce the number of parameters of the model by the two characteristics of local connection and weight sharing,so that the model can achieve a very deep level.Deep learning can directly learn and extract relevant features from a large quantity of image data,so as to reduce the complexity of the modeling process.In this paper,we study the method of image segmentation based on deep learning.The method of semantic segmentation by deep learning is using pixel level annotated image and upsampling layers,deconvolutional layers and other special layers to restore features that extracted from common convolutional networks to the size of the original image,so as to achieve an end-to-end learning.In this paper,we present a new network architecture called Fast-Seg Net based on the existing networks.The new architecture has achieved the depth of 28 layers by using the residual network structure,dilated convolutions,asymmetric convoluitons,and batch normalization.The precision has been improved,at the same time,the running speed has been kept.The training speed of the network has been improved by the method of transfer learning.In addition,this paper presents a set of tools and theprocess of the pixel level images labeling.The software of the pixel level image labeling inspection tools was written.Photos of street scene in the campus of Jilin University was collected.We labeled 1566 qualified images,and trained the Fast-Seg Net by using them.We have achieved good results that the m Io U reached 70.2%.