Research On Image Semantic Segmentation Of Road Scene Based On Convolutional Neural Network

Image semantic segmentation is a kind of important technology that uses the high-level semantic features in the image to predict,recognize and classify the semantics of a single image pixel.With the continuous in-depth development of modern artificial intelligence technology,in the research and practice of the frontier field of computer vision,image semantic segmentation has become the most basic part of the research.It has increasingly profound application and development value in many related fields at home and abroad,such as unmanned driving cars,medical image analysis,photography and so on.The early image segmentation methods have poor segmentation performance and can not meet the needs of semantic segmentation in complex urban road scenes.With the continuous development of deep learning technology,the traditional semantic segmentation has been promoted to pixel level semantic segmentation.The semantic segmentation method based on convolutional neural network has gradually replaced the traditional semantic segmentation method.Therefore,pixel level segmentation of image semantics based on convolutional neural network has also become a hot research at present.The application of image semantic segmentation in road scene is mainly divided into two kinds: one is to segment different objects in road scene;The other is to segment different objects of a certain kind of objects more finely.The semantic segmentation of road scene studied in this paper is to realize the multiple segmentation of different types of objects in urban road scene based on convolutional neural network.The convolutional neural network model is complex,with a large number of parameters,and is vulnerable to other noises.Therefore,semantic segmentation of road scene is challenging.Therefore,in view of the above problems,the specific research contents of this paper are as follows:(1)Due to the complexity of Fast-SCNN model and large amount of parameters,the speed needs to be improved during calculation,and there may be some redundant information.Therefore,it is necessary to find an effective evaluation method to judge the importance of parameters and cut off some unimportant connections,so as to reduce the redundancy of the model and improve the reasoning speed of the model.In this paper,model pruning is proposed in the global feature extraction part,Fine tune this part to reduce the amount of parameters and calculation.Experiments show that this method is effective and can improve the reasoning speed.At the same time,receptive field is also one of the reasons affecting the segmentation effect.When Fast-SCNN model performs semantic segmentation,it has the problems of complex background of similar targets and difficult to distinguish different categories of targets.Therefore,multi-scale feature information fusion is introduced to fuse the feature information of different sizes with the high-level semantic feature information,which can identify more shallow feature information,increase the size of receptive field of the network and improve the accuracy of segmentation.Experiments on cityscapes data set verify the effectiveness of this method.(2)When image information is restored in the decoder part of fast SCNN network,due to the problems of network structure and segmentation method,the image information is compressed excessively,and a lot of feature information will be lost in the process of image information restoration,resulting in low segmentation efficiency.Therefore,in the decoder part,expansion decoding is adopted,and three expansion convolutions are densely connected together to obtain multi-scale information,More abundant and detailed feature information is obtained,which effectively improves the segmentation accuracy of the network model.Experiments on cityscapes dataset show that the accuracy of the improved method is higher than that of the original model.(3)In the process of image semantic segmentation,it will be disturbed by other noise.In this process,the recognition mechanism will pay attention to some useless information,resulting in the loss of some useful feature information.To solve this problem,this paper introduces two different attention mechanisms to use the feature information obtained from channel and space to enhance the network’s further attention and learning of target detail information,The interference caused by other noise information is reduced to a certain extent.It can pay attention to more useful detail information in the rich image feature information.At the same time,experiments are carried out on cityscapes data set and ade20 k data set to verify the effectiveness of this method.