Conditional Bidirectional Learn And Inference Based On Wasserstein Distance

Image generation technology has always been an important research direction in professional fields such as computer vision and computer graphics,and is widely used in industry.After the dedication of many scientists,the results of the deep learning image generation are still unsatisfactory.The main difficulties and challenges are the diversity,authenticity,stability,and controllability of the image generation results.In recent years,due to the advent of generative adversarial networks,image generation technology has made great progress in many aspects.However,due to the disadvantages of generative adversarial networks,such as unstable training,inability to judge convergence,and mode collapse,Therefore,there is still room for development of image generation technology.This thesis proposes a Wasserstein Bidirectional Learned Inference(WBLI)model based on Wasserstein distance,which integrates the encoder and generator bidirectionally into the Wasserstein generative adversarial network to solve the problem of mode collapse and training instability in generative adversarial networks,Thereby enhancing the learning ability of the generated model.The coding network maps the data space to the hidden variable space,and then uses the generator to map the hidden variable space back to the data space.Because the samples in the data space distribution are diverse,the generated samples obtained through this bidirectional model are also diverse.The problem of mode collapse in generative adversarial networks is greatly alleviated.The thesis proposes to integrate the conditional constraint information into the WBLI model(CWBLI),integrate the classifier into the WBLI model to solve the uncontrollability of the WBLI generated samples.The loss function of the classifier is further integrated into the generator by weighting with the important weights proposed in the thesis,and the Learning direction of the model is automatically adjusted at different stages of model training,thereby ensuring the stability of the gradient descent direction when many network structures are trained simultaneously.The experimental results show that the CWBLI model alleviates the existing problems in the image generation technology and the generation adversarial network to varying degrees,and generates generation samples that meet the data set standards.