Application Of Generative Adversarial Networks In Image Delta Generation And Comic Line Coloring

In recent years,with the introduction of Generative Adversarial Networks(GAN),Generative Adversarial Networks have been widely used in image generation,image conversion,and other fields.Aiming at the problem that the contrast models are not effective in image data generation and comic line coloring tasks,this thesis proposes an improved model architecture based on generative adversarial networks and obtains generated images with considerable effect through training.The results are as follows:(1)Aiming at the problems of mode collapse,gradient disappearance,and training instability in contrast image data generation models during training,an image data generation model based on the generative adversarial network is proposed.The model generates the desired image through a given noise vector.In the experiment,the model adopts a generator and a discriminator composed of a branch residual network,which can provide the network with receptive fields of multiple scales and improve the representation ability of the network.Aiming at the problem that mainstream generative adversarial networks can only generate a single type of image,a model that can generate multiple types of image data is proposed.The generator of the model consists of a network with multiple branch outputs.It can generate corresponding types of image data by increasing or decreasing the number of branches of the network model.To better guide the network for training,a dynamic learning rate is designed,which can dynamically adjust the decay of the learning rate according to the specific task.The experimental results show that the image data generation model proposed in this thesis can generate image data with richer diversity,higher definition,and better quality than other contrast models.(2)Aiming at the problems of inaccurate coloring,low contrast,and poor quality of the colored images generated by the contrast comic line coloring model,a comic line coloring model based on the generative adversarial network is proposed.The model generates a comic-colored image based on the given comic line image.In the experiment,the model adopts the generator of a U-shaped network structure and combines the branch residual network with the Markov discriminator to design the discriminator of the model.To perform better feature extraction,an improved channel space fusion attention network model is proposed to strengthen the important information in the feature map and suppress the non-critical information.According to the U-shaped structure of the generator,the attention module is embedded in the network connections of multiple scales,so that the attention module can process the multi-scale image features of the network model.To be able to generate higher-quality colored maps,adversarial loss,minimum absolute value error loss,and perceptual loss are used to constrain network training,which can better guide model training.The experimental results show that the comic line coloring model proposed in this thesis can generate more accurate coloring,higher contrast,and better quality comic line coloring images than other contrast models.