Research On Intelligent Water Affairs Oriented Image Generative Model

In recent years,with the rapid development of artificial intelligence,image stylization technology is playing an increasingly important role in industrial production,and is also widely used in abstract art creation,animation simulation,portrait generation,and other fields.Image stylization technology provides great convenience for improving the efficiency of scientific research and intelligent industrial production.However,in actual project development,image stylization techniques also face some problems: the existing stylization techniques do not provide high feature extraction for content and style images;The efficiency of model generated images is not high;Problems such as poor quality of generated images that are relatively difficult to solve.It has seriously affected the research and application of stylization technology.Therefore,improving the quality of stylized images and maintaining the same structure of the generated images is of great significance.In the context of the development of smart cities,smart water has also emerged.It improves the management level of urban water systems by using technical means such as selfcontrol,perception,and intelligence.Use cameras to collect water meter images,reservoir images,river channel images,and other data,and use artificial intelligence to identify and upload readings to the management platform.To solve the problem of requiring diverse training data and difficult collection in recognition models,image stylization technology is used to assist in expanding the training dataset.In order to solve the problems of data shortage and incomplete data types,this paper studies image stylization methods for intelligent water services.From the aspects of image generation quality and stylization accuracy,it studies the improvement of image generation quality and the maintenance of image structure,so that the generated image can be consistent with the color and structure of the original image,while possessing the style characteristics of the style image,thereby truly realizing a variety of water meters The expansion of training data such as reservoirs meets the requirements for intelligent water training datasets.The main research contents of this article are as follows:Aiming at the problem of poor image quality and color deviation in generated images existing in Cycle GAN,an image quality based adversarial neural network model(IQCycle GAN)was proposed.Based on the existing Cycle GAN network,this algorithm replaces the transcoding block and decoding block in the generated network with the RRDB residual structure.At the same time,we will use color loss and perceptual loss to correct the color deviation of the generated image.The experiment shows that the problem of low image quality generated by Cycle GAN is solved.To solve the problem of global structure changes and poor local style feature migration in existing images generated by Cycle GAN,a structure preserving Cycle GAN(Str PCycle GAN)based adversarial neural network model was proposed.Based on the existing Cycle GAN network,this algorithm uses a U-net network with a residual structure based on the self attention mechanism as the generator of Cycle GAN.A network that combines the residual structure with a U-net structure for detail feature extraction will make the detailed features of the generated image more obvious.Experiments show that this algorithm solves the loss of global structural feature information and detailed style feature information in the generated images by Cycle GAN.To solve the problem of poor image quality and structure retention,a structure retention generation countermeasure network model combining image quality was proposed.Combine the designed IQCycle GAN generator with the designed Str PCycle GAN generator,and add a self attention mechanism to ensure the global characteristics of the image.In terms of the loss function,introduce a color loss function and a style perception loss function,while adding a smoothing loss function to help improve the quality of image generation.Experiments show that the proposed model can generate high-quality structure preserved images.The effectiveness of this method is verified.