Research On Image Semantic Segmentation Algorithm Based On Deep Learning

In recent years,owing to the improvement of big data and the computing ability of hardware,the rapid development of deep learning in artificial intelligence has become one of the hot topics in the field of artificial intelligence.As an important branch of artificial intelligence,the research of computer vision has also begun to use the deep learning methods and has gained many outstanding achievements.Deep learning has made a breakthrough in image recognition,target detection,image retrieval,image content analysis,object tracking and so on,making computer vision applications begin to enter into people's life.Because of the high requirement of computer vision task in industry,the research of visual task based on deep learning is still a hot issue.Semantic segmentation is an important research problem which is a premise of the field of computer vision.The performance of subsequent algorithms is affected directly by the advantages and disadvantages and segment accuracy of image semantic segmentation algorithm.Therefore,the researches of sematic segmentation algorithm are available.With the development of deep learning which is penetrated to other research filed,the research of image semantic segmentation algorithm based on deep learning has become hot issue.Therefore,the research of image semantic segmentation algorithm has entered a new stage of development.Graph theory model was built by traditional image semantic segmentation algorithm according to the low-level visual information of the image pixel itself,then classify pixels use this built model.However,the image segmentation with clutter background is ineffective and requires human intervention.Due to Convolution Neural Network can learn and extract abundant features from image,trains the network using a lot of labeled specimen.So,it can get the mapping between image and image semantic label.Although,semantic segmentation algorithm based on deep learning has good performance,how to promote the accuracy of semantic segmentation is still hot issue.To solve the problem of how to promote the accuracy of image semantic segmentation through image features information.Therefore,image semantic segmentation algorithm based on deep learning was discussed in this thesis.Firstly,this thesis zoom image to three different scales,and feed it to a separate feature extraction network,extract features through training.This method can increase the number of origin data sets,and can extract image features from three dimensions.Fully convolution network was used in network structure of this thesis,fuse the three different scale feature maps,and output the segmentation results,use the standard data sets to validity the semantic segmentation algorithm.Then,an improved encoding and decoding network is proposed in this thesis.Encoding phase features are extracted from the image and the Decoding phase get the Segmentation image which has same size with origin image with up-sample method based on location information recorded by down-sample method.The convolution was replaced by the network of Inception model in the encoding processing.Inception model extracts feature information of different receptive fields,with the different kernel size.Finally,we evaluated our method in CamVid data sets,experiment result show that the accuracy of image semantic Segmentation have improved in some degree.