Research On Image Semantic Segmentation Method Based On Deep Convolution Neural Network

Image semantic segmentation is widely used in many area,such as autopilot,scene understanding,medicine and so on.However,the existing semantic segmentation techniques have many problems such as rough edges,low segmentation accuracy,tedious label process The generalization ability of the segmentation model is low and can only be used to a specific environment.So a feedback weak supervised semantic segmentation method for deep antagonistic neural networks and transductive transfer learning method for deep adversarial neural networks are proposed based on the two-level merged upsampling method.The segmentation accuracy is improved,the label process is simplified,and the generalization ability of the network is improved.The research in this paper is of practical significance to the wide application of semantic segmentation technology.Firstly,a semantic segmentation method based on two-level merged and upsampling is proposed to solve the problem of segmentation.The method of depth deconvolution is used to replace the bilinear interpolation upsampling method in semantic segmentation networks.By learning multiple sets of upsampling convolutional filters and preserving the maximum weight relative position index of maximum pooling,the segmentation network combines the learned information in the convolution filter with the retained maximum weight position index.The depth antagonistic neural network is used to distinguish the output prediction results of the segmentation network,so that the lost detail information can be compensated and the edge precision of segmentation can be improved.Secondly,a feedback weakly supervised semantic segmentation method based on depth adversarial neural network is proposed to solve the tedious problem of label process.The training data set is divided into two parts: labeled data and unlabeled data.The labeled data is used to supervised training and the unlabeled data is used to weakly supervised learning.By using the discriminant mechanism of depth adversarial neural network,the output results of supervised training network model and the output results of weak supervised training network,output results are used to carry on the adversarial training.It reduces the demand for labeled data and ensures high accuracy of segmented networks.Finally,aiming at the problem of low generalization ability of segmentation model in specific environment,a transductive transfer learning semantic segmentation method is proposed.The semantic segmentation based on two-level merged and upsampling is used as baseline network.The source data and the target data are extracted by convolutional filter in baseline network at the same time.Then,the source images and the target images are divided into spatial regions,and the problem to spatial distribution alignment between the two domains is improved by using the inherent spatial structure in the urban scene image.Next,the high-level feature output features in the source domain are mapped to the shallow feature space in the target domain and adversarial training is performed to reduce the difference between the source domain and the target domain in the feature space.By transfering the network of the source domain to the target domain,the segmentation model obtained from the training of the source data can get a better segmentation effect in the target domain,which improves the accuracy of the segmentation prediction of the target domain data in the segmentation network.