Abstract:Aiming at the translation between different scene images, we propose an improved generative adversarial network model that can generate high-quality target scene images. In the process of generating a target image, the spatial position information of the original image will lose due to down sampling. Therefore, a generative network that includes jump connections and residual blocks is designed in this paper. By adding multiple jump connections to the network, we can keep the spatial position information of the image transmitting in the network. At the same time, to improve the stability of the generated image during the training, we introduce the Structural Similarity Index Measure (SSIM) as a structure reconstruction loss to guide the model to generate a target image with a better structure. In addition, in order to make the translated target scene image retain more color details, we add an identity preservation loss, obviously enhancing the color expressiveness of the target generated image. The experimental results show that the improved generative adversarial network model proposed in this study can be effectively applied in scene image translation.