Research On Semantic Segmentation Of Deep Convolutional Neural Network

Recently,artificial intelligence has been developing rapidly.As one of the most popular technology,Deep Learning is widely used in many fields and plays a key role in them,such as intelligent robot,speech recognition,image processing and so on.As a core technology,Deep Learning is also applied to image semantic segmentation.Compared with the traditional semantic segmentation technology,deep-learning segmentation algorithm does not need any artificially designed features and does a good job at extracting the in-depth features.Therefore,it has been favored by many researchers.In the deep-learning semantic segmentation algorithm,neural network is the most important part.In general,the more layers in network,the stronger ability to extract features.However,with the increase of levels,the features extracted in the network will become more abstract,and more and more details in it will be lost.So this paper proposes several effective methods to optimize one of the classic networks(Deep Lab-LargeFOV).The effectiveness of these methods is proved by several experiments.The main contribution of this paper is introduced as follow.Firstly,the ideas of share net are used to optimize the in-depth features in the network.Then,we explore and optimize a variety of feature-fusion models which are in shared ways.Finally the best optimization scheme is obtained experimentallySecondly,global context optimization is applied to the shared,feature-fusion model.We explore many globally-feature-fusion methods and improve some of the existing globally-feature-fusion methods.Finally the best way to combine with global features is found.This optimization can make the in-depth features more expressive.Thirdly,we propose a cross-layer,feature-fusion method based on the middle-layer features,which optimizes the output of up-sample layer.The innovation of this method is that it is mainly a feature-level fusion based on middle-layer features,rather than a simple fusion of heat maps.Fourthly,we propose a local,similarity-constrained algorithm based on shallow-layer features,which optimizes the segmentation ability of the network by similarity constraint of shallow-layer features.In our experiments,we find out the most suitable shallow-layer features as well as the best way to extract local-correlation features,which can further increase the effect of our method.By comparison test between original model and our optimized model,it is proved that our methods can increase the semantic-segmentation effect of normal convolutional neural network.Furthermore,our optimization is also superior to the general method based on fully-connected conditional random fields.