Few-Shot Image Generation Based On Generative Adversarial Networks

In recent years,the deep learning models are growing fast and have been widely used in various fields.However,most neural networks now still rely on a large number of data for training;especially for the generation task,the scarce training data will seriously reduce the quality of image generation,and the few-shot learning methods can alleviate this problem to a certain extent.The few-shot image generation task aims to utilize a small number of samples from the target class to synthesize rich images.This paper will focus on the few-shot image generation task,complete the task definition and survey the related works at home and abroad.Then it will focus on two tasks: the few-shot image generation under the traditional definition and the few-shot image generation based on knowledge transfer,and point out the disadvantages of the current methods.In the first task,the current fusion-based solutions consider introducing reference image features in every layer of decoder during generation,which will aggravate the feature misalignment and cause the generation of artifacts.In the second task,the current methods solve the problem by predicting the distribution of target data to achieve knowledge transfer or by applying data augmentation to the target data.The above two methods are difficult to achieve satisfactory results when the data is scarce.To solve the above problems,this paper proposes new algorithms for the two tasks from model design and deep learning strategy:The work of this paper mainly includes the following three points:1)For the few-shot image generation task under the traditional definition,this paper proposes a few-shot image generation algorithm GLFGAN based on local fusion and global fusion procedures.Through the local fusion block,the algorithm realizes the fusion of the local features from basic image and the features from reference image.Through the non-local attention mechanism,the global fusion block achieves the alignment within local features to obtain more natural global features.The local and global discriminators guarantee the authenticity of the generated images from multiple levels.2)For the few-shot image generation task based on knowledge transfer,after defining the conditional few-shot image generation task,this paper proposes the idea of knowledge evolution,and designs the evolution strategy based on the pretrained network VQGAN.In the introduction of evolution location,through the generalization experiment,this paper finds out the modules that need to participate in the evolution.Then it adds a new embedding layer to process the new category labels.In the introduction of evolution method,this paper proposes the method of incorporation initialization to initialize the parameters of the new module,and obtains the network parameters that need to be evolved through correlation analysis.3)Through measuring the FID,LPIPS and KMMD values on extensive experiments,this paper verifies the superiority of our methods compared with other methods in terms of the rationality of the generated content,the authenticity and richness of the generated content.