Research On Image Semantic Segmentation Based On Convolutional Neural Network

Image semantic segmentation has always been a hot topic in the field of computer vision.Because of its powerful learning ability,convolutional neural network has achieved remarkable results in semantic segmentation tasks.This paper makes an in-depth study and analysis of the existing image semantic segmentation algorithms based on the convolutional neural network,and finds that the comprehensive utilization of the feature information in the deep and shallow layers of the convolutional network can significantly improve the segmentation effect.Therefore,from the perspective of effective use of shallow level visual features and deep level semantic features,this paper proposes an image semantic segmentation method integrating multi-level feature information with the purpose of improving the quality of image segmentation by using the hierarchical feature map of convolutional neural network.For this method,the following research work has been carried out:1、Based on DeepLab v3-plus network model,the void space pyramid pooling module is transformed.The method of extracting the same feature map from multiple branches of the module is changed into the form of extracting multi-level feature map from multiple branches to form its own multi-level feature fusion module.Therefore,the parallel extraction and fusion of multi-scale feature information is realized.And this paper adds global average pooling module to supplement more global context information for multi-level feature fusion module.Therefore,the global feature of the model can be optimized,and the recognition ability of the model to the target object can be improved,and the global expression ability of the deep feature to the image can be strengthened.2、In this paper,the cross-layer feature fusion method is adopted to continuously iterate the deep features to enrich the low-level visual information,so as to generate denser feature maps in each layer path.The deep and shallow feature information stimulate each other in the learning process by iteration,so that the information beneficial to segmentation in the feature map can be used more effectively.And this paper explores the effects of different expansion rates,different channel Numbers of feature graphs,different global average pooling operations,and different cross-layer feature fusion methods on the segmentation results.Finding the most suitable scheme can determine the optimal segmentation ability of the model structure.