Image Scene Transformation Based on Dynamic Mutual Information Generative Adversarial Networks

Generative Adversarial Network (GAN) has outstanding performance in the field of image scene conversion. It has been applied to visual tasks such as the conversion of satellite remote sensing images to floor plans, and has important application value in industrial production and other fields. However, the long training time of GAN model has become a common phenomenon in the field of image generation. Increasing the iteration speed can effectively speed up the project development progress, so speeding up the model training speed has gradually become a research hotspot. This work uses feature matching and maximizing mutual information, assigns dynamic weights to data such as features and information, maximizes the loss of distribution difference, speeds up the training cycle, dynamically adjusts the image data input according to the model learning progress, and makes the generated image retain more details. A dynamic mutual information generative adversarial network model is proposed. The experimental results show that compared with other models, the amplitude and duration of this model are reduced by 7.1% and 4.1%, respectively, which can reduce the training cost.