Research On Realistic Image Generation Algorithm Based On Generative Adversarial Network

With the rapid increase of the users of social media worldwide,there is a higher demand for image data processing technology,text image technology has become a hot issue in the field of computer vision synthesis.Because of the semantic gap and heterogeneity gap between text data and image data,how to obtain the effective features of text data and how to construct the mapping between text semantic space and image semantic image has become the core problem of text generated image research.Due to the game theory of generative adversarial network and the excellent performance of deep neural network in processing text and image data,image generation based on generative adversarial network has become the mainstream method in the field of computer vision synthesis.It makes use of the game between generator and discriminator to make generator generate realistic images which can be distinguished from the real ones during training.However,the existing image generation models based on generative adversarial network still have some shortcomings,such as it is not easy for discriminators to measure semantic similarity between different modal data in the training process and the limitation of using attention model in generator to generate fine-grained images in matching text semantics.In view of the above mentioned problems,this paper proposed corresponding solutions respectively,and conducted a large number of comparative experiments with various mainstream methods on open source data sets.The main research work and innovations of this paper are as follows:1.Aiming at the problem that it is difficult to measure the similarity of semantic matching between the generated image and text in the training process of discriminator,a text retrieval algorithm based on semantic matching is proposed.On the basis of preserving the traditional method of calculating semantic similarity by using likelihood logarithm,the Jaccard Distance is introduced to reflect the content similarity between text and image,and the content loss function is calculated in Hamming space.At the same time,the convolutional neural network is used to learn image feature hash function and text feature hash function.The prediction model and label alignment processing of the generated image are constructed.Experiments on open source data sets show that the proposed algorithm can effectively improve the ability of discriminator to generate images due to the traditional hashing retrieval method.2.Aiming at the problem of missing semantic feature information in matching text semantics of fine-grained images generated by using attention model in generator,an algorithm is proposed to solve the attribution of attention score obtained by attention mechanism.Predictive learning is introduced to mine the relationship between the feature attention score matrix obtained from the text and the local image results that conform to the text semantics,and the quality of the image generated by the generator is optimized.Experiments on open source data sets show that the proposed algorithm is superior to existing image generation algorithms.