Research On Image Semantic Segmentation Method Based On Generative Adversarial Network

Image Semantic segmentation of image is a very important task in computer vision.It achieves object segmentation at pixel level by predicting and classifying every pixel in the image.This technology is currently used in a wide range of areas of people's lives,such as medical imaging,autonomous driving,and geographic information systems.Accurate segmentation of images is the basis of deep understanding of images,so it has great application value for the research of image semantic segmentation tasks.In this paper,the image semantic segmentation method based on Generative Adversarial Networks is deeply studied,and two different image semantic segmentation methods are implemented from the perspective of integrating multi-scale information of image and increasing the number of image feature extraction.The main work of this paper has thefollowing three points:(1)Starting from the semantic segmentation model of encoder-decoder structure,combined with conditional Generative Adversarial Networks,a Segmentation-Adversarial Networks model is designed,and the semantic segmentation network and discriminant network are combined.The segmentation network is designed to output segmentation images and discriminate.By comparing the difference between the segmented image and the real marker image,the network guides the segmentation network optimization model parameters to further improve the segmentation accuracy,and the two networks alternately train to form a confrontation.Experiments show that the segmentation-against semantic model further improves the segmentation effect of the segmentation network.Image Semantic segmentation of image is a very important task in computer vision.It achieves object segmentation at pixel level by predicting and classifying every pixel in the image.This technology is currently used in a wide range of areas of people's lives,such as medical imaging,autonomous driving,and geographic information systems.Accurate segmentation of images is the basis of deep understanding of images,so it has great application value for the research of image semantic segmentation tasks.In this paper,the image semantic segmentation method based on Generative Adversarial Networks is deeply studied,and two different image semantic segmentation methods are implemented from the perspective of integrating multi-scale information of image and increasing the number of image feature extraction.The main work of this paper has the following three points:(1)Starting from the semantic segmentation model of encoder-decoder structure,combined with conditional Generative Adversarial Networks,a Segmentation-Adversarial Networks model is designed,and the semantic segmentation network and discriminant network are combined.The segmentation network is designed to output segmentation images and discriminate.By comparing the difference between the segmented image and the real marker image,the network guides the segmentation network optimization model parameters to further improve the segmentation accuracy,and the two networks alternately train to form a confrontation.Experiments show that the segmentation-against semantic model further improves the segmentation effect of the segmentation network.(1)Using the idea of generating anti-network to optimize the image semantic segmentation task,using the semantic segmentation network as the generator to realize the preliminary segmentation of the image,and then training a discriminator to judge whether the segmentation result is from the generator image or from the real Label images,segment the network and discriminate the network for confrontation training.Applying confrontation learning to the semantic segmentation task of the image,the gap between the segmented image and the real image can be well discriminated and fed back into the semantic segmentation network,thereby further improving the segmentation effect.(2)A method of image semantic segmentation based on USGAN is proposed.Firstly,the encoder-decoder semantic segmentation network USegNet is proposed.The coding stage uses convolution and pooling operations to extract the feature information of the image.The decoding stage is not only adopted.The maximum pooling improves the boundary partitioning and reduces the training parameters.It also adds a jump connection between the encoder and the decoder,and combines the shallow features and the deep features of the image to extract multi-scale information.As a segmentation network,USegNet forms a Segmentation-Adversarial Networks with the discriminator network to train the network in a confrontational manner.(3)A RUGAN based image semantic segmentation method is proposed.The shape of U-Net network structure and residual block residual block idea are used to propose another encoder-decoder semantic segmentation network ResU-Net.The network can extract the detailed features of the image to the greatest extent and can cascade the features of different levels,and the feature expression of the image is rich and accurate.On this basis,the discriminator network is introduced to further optimize the segmentation effect,and a Segmentation-Adversarial Networks model based on RUG AN is formed.The research on the two image semantic segmentation methods mentioned in the paper achieves the expected purpose by conducting experiments on the public dataset.The experimental results show that the image semantic segmentation method based on Generative Adversarial Networks has a good improvement in accuracy.