Research Of Texture Filtering Algorithm Based On Generative Adversarial Network

Texture filtering is a low-level image processing,its main operation is to preserve the prominent structure and remove the oscillating texture.The key to this problem is to find the right boundaries to separate structure and texture.While The traditional filtering methods spend a lot of time on the hyper-parameter adjustment,deep learning methods require high-quality datasets.In this paper,we propose two schemes to study the GAN in texture filtering.In the first task,a semi-supervised texture filtering method is proposed to train GAN using a small count of labeled data and a large amount of unlabeled data.In the training stage,the loss function is designed for labeled and unlabeled dataset respectively.The information extracted from the shallow feature layer and the deep feature layer of the pretrain network should be used reasonably.Shallow features preserve edges,while deep features recognize semantic content and eliminate small scale texture variations.This contribution effectively addresses the main challenge of texture filtering,which is to distinguish structural content from unstructured textures at the pixel level.The extracted information improves content and color consistency between the original image and the filtered image,especially for unlabeled samples.This scheme has two advantages:using a large amount of unlabeled data to reduce the over-fitting of labeled data and improve the generalization ability of the model.In the second task,mixed-scale dataset are used to guide GAN to obtain high quality filtering images.The purpose of using mixed-scale data sets is to reduce the harsh requirements of data preparation and to lower the accuracy requirements of large and small scale data.By using the prior information of edge detection,GAN can distinguish semantic structure edge and internal texture.Mixed-scale data sets provide the flexibility for GAN training to filter pixels at different locations:small scale data is used for pixels on/around the edge of the structure,while large scale data is used for pixels within the textured region.Finally,based on qualitative and quantitative evaluation,the scheme has obtained advanced filtration results,especially in the aspect of edge retention quality.Both schemes can achieve the performance of the existing methods and reduce the need to determine the optimal parameter values.The two schemes presented in this paper greatly reduce the time and effort required to reconstruct the marked dataset,especially for pixel-level fine operations while demonstrating the application of texture filtering to assist in various tasks such as detail enhancement,image abstraction,pencil drawing,super-pixel and edge detection.