Research On New Style Migration Network For Sample Expansion

In recent years,with the rapid development of computer vision,semantic segmentation technology plays a more and more important role in industrial production.At the same time,it is also widely used in various fields of daily life.Semantic segmentation technology provides the possibility for enterprises to improve production efficiency and realize the intellectualization of industrial production in order to obtain higher competitiveness in the market competition.However,while semantic segmentation network is widely used,it also faces some difficult problems.The fully supervised semantic segmentation network needs to manually label the labels corresponding to the samples during training,which brings high labor and time costs,which seriously restricts the development and application of semantic segmentation network.Therefore,reducing the cost of manual annotation while improving the effect of semantic segmentation is of great significance to the rapid deployment and application of semantic segmentation network.Aiming at the problems of difficult label making and manual labeling in the segmentation of graphite electrode steel seal characters in specific industrial scenes and the segmentation of urban street scenes in automatic driving technology,this paper proposes a new style migration network Cycle GAN-AD for sample expansion.From the perspective of data samples,the network reduces manual annotation and improves the accuracy of semantic segmentation,which is of great significance for the wide application of semantic segmentation network in the future.From the perspective of sample expansion,the simulated pictures with their own labels obtained from style migration can be combined with the real pictures to construct the data set,so as to improve the segmentation effect of the fully supervised semantic segmentation network.Even if there are only a few real picture labels,the training with the pictures generated by style migration of simulated pictures can ensure the accuracy of semantic segmentation;From the perspective of unsupervised semantic segmentation,the self labeled simulated pictures obtained from style migration can be used to construct a separate data set to realize unsupervised semantic segmentation.The main research work is summarized as follows:(1)A sample expansion method based on image style migration is proposed.The computer modeling image with its own label is used as the source domain data,and the real image collected by the camera is used as the target domain data.Through the style transfer training of circularly generated countermeasure network,the computer modeling image style is transferred to a simulated image close to the real image distribution.The simulated image with label can be used as the training sample of fully supervised semantic segmentation network,and the sample expansion is realized.(2)An improved style migration network Cycle GAN-AD is implemented.Based on Cycle GAN,a channel attention mechanism is added to the generator.The channel attention mechanism is mainly composed of maximum pooling layer,average pooling layer and full convolution network layer.In the process of network training,the style migration result is improved by extracting the effective information in the picture.Dense connected convolution network Dense Net is used to replace residual network Res Net in Cycle GAN.In Dense Net,each layer is composed of the previous layers.The connection between the network layers is closer to ensure that the details of the original picture are not damaged due to the deep convolution network layer.The improved discriminator is a Markov discriminator to improve the discrimination ability of the discriminator.Experiments were conducted in three data sets: Horse migration into zebra,simulated sample migration into graphite electrode,and GTA5 game image migration into urban street view.Verify that the image quality generated by the improved network is better than Cycle GAN.(3)Semantic segmentation experiments are carried out in graphite electrode and urban street view data sets.U-Net and Deep Lab V3 + are selected as semantic segmentation networks.The image generated by simulated image style transfer can be used as semantic segmentation data set for experiment alone,or can be used as training set together with real sample data set to realize sample expansion experiment.The real image test set is tested to analyze the impact of the image generated by style transfer on the semantic segmentation results.The experimental results show that after the image generated by Cycle GAN-AD network is expanded for samples,the segmentation effect is significantly improved.The MIo U value of graphite electrode steel printed characters is up to 82.60,and the MIo U value of urban street view is up to 46.90.It can be seen that the sample expansion method proposed in this topic is expected to significantly reduce the workload of manual annotation and obtain high-quality training samples,which is more conducive to the popularization and application of semantic segmentation and other algorithms in the process of industrial production and life and improve efficiency.