Text To Image Generation Based On Generative Adversarial Network

With the development of deep learning,especially the proposal of Generative Adversarial Networks,the task of generating images from text has become an important research direction in the field of image generation.The text-to-image task requires a text description and generates image results that conform to the semantics of the text.In the text-to-image generation methods,there are challenges such as unstable network training,insufficient authenticity and diversity of generated image details.In order to improve the authenticity and diversity of image generation,this paper has done the following work:First,to achieve stable fine-grained text-to-image generation,this paper proposes attention and convolution-normalized generative adversarial networks.The method achieves image generation by introducing a stacked generative adversarial network with attention mechanism,and adds convolution normalization mechanism to the up-sampling module of the discriminator.High-resolution fine-grained image generation is achieved step by step through multiple stage generators,and the model training process is stabilized through convolution normalization,which improves the quality of image generation.Secondly,in view of the problem that the networks of the text-to-image generation methods are too complex and lack attention to the important information of the text,this paper proposes an attribute-based partition condition fusion text generation image method.A single-stage generative adversarial network is constructed through a deep text-image fusion module,and an attribute partition fusion module is proposed.The attribute partition fusion module first extracts the attribute words in the text,and then partitions the image features according to the matching degree between the image sub-region and the attribute words,and uses the corresponding attribute word vectors for each partition to perform conditional fusion to enhance the deep fusion of image and attribute conditions.The model improves the variety and detail authenticity of the generated images.Finally,for complex scene description texts with multiple objects,this paper proposes a deep fusion generative adversarial network based on object semantics.The generator of the model consists of two parallel network branches,the global image generation branch and the local object generation branch,respectively.The global branch takes the entire text sentence as constraints,and the local branch takes object labels and bounding boxes as constraints.The discriminator of the model includes global image and local object feature extraction branches,and combines the features of the two branches for discrimination.The model is composed of a deep text-image fusion module,which realizes single-stage complex scene generation and improves the generation quality of object features.