Image Generation Of Generative Adversarial Network Based On Channel Feature Learning

The generative Adversarial network relies on its unsupervised learning method and powerful generative ability,and it has attracted people’s attention since it was proposed in 2014.While GAN is constantly proposing new models,it also encounters many new problems.Among them,the long-distance dependency fitting ability is weak,the image global features are inconsistent,and the resulting poor quality of the generated pictures are constantly appearing,which limits the further application of GAN.This thesis focuses on the research of the generation method of generating adversarial network image based on the channel feature learning,and proposes a method to effectively use the channel feature to improve the quality of the generated image.By tracking the latest developments in related fields at home and abroad,this thesis also combines the attention mechanism to further improve the network,enhancing the network’s ability to fit long-distance dependencies.The specific results are as follows:1.This thesis proposes a generative Adversarial network model based on channel fea-ture learning,which has achieved good results in GAN image generation tasks.The model uses WGAN as the basic network,uses Wasserstein distance to calculate the gap between the sample and the real distribution,adds a feature discrimination mod-ule to the feature layer to improve the quality of feature generation,and implements gradient penalties across the entire network to make the generation of each network Layer convolution has a clear learning goal,which improves the quality of the fea-tures generated by the generative network during the generation process,thereby improving the final image generation quality.2.Aiming at the problem that the image generation method based on channel feature learning relies on the lack of long-distance fitting ability,this thesis proposes the method of adding mixed attention to improve,so that the feature distribution of the network is more reasonable.The image features are first enhanced by the attention mechanism to the key features,and then corrected by the feature discriminating network,which further improves the long-distance dependence and fitting ability of the network.Finally,a large number of experiments have proved the effectiveness of the method proposed in this thesis.3.This article also proposes a transformer-based discriminative network model,and compares the performance of different discriminant networks of Transformer and hybrid attention mechanism in tasks,and also proposes an effective training method to balance the training level of Transformer and generation network.And the fea-sibility of this method is verified in the experiment.