An Ancient Calligraphy And Painting Image Super-Resolution Method Based On Improved Generative Adversarial Network

Image super-resolution reconstruction aims to restore one or more low-resolution images to high-resolution images by certain means.As images have become the main medium for communication in human production and life,improving the resolution of images is in line with the pursuit of modern public and is of practical research significance.Due to the rapid progress of computer science,the research on deep learning technology and image super-resolution fusion has also made great development,but there are problems such as insufficient representation of high frequency details and smooth texture details in the reconstructed image results of existing methods.In this paper,under the goal of improving image perceptual quality,the following main works are done.(1)Based on the image super-resolution method with the architecture of generative adversarial network,the generator structure is improved by using a residual dense block structure incorporating the channel attention mechanism,which adaptively adjusts the features using the correlation between channels and improves the discrimination ability of the network for effective feature channels;a convolutional layer with a kernel of 1 × 1 is used in the basic dense block to effectively reduce the modeling parametric and training The use of a 1×1 convolutional layer in the basic residual cascade block effectively reduces the covariance and training complexity of modeling and avoids overfitting;the use of techniques such as spectral normalization during training stabilizes the model and alleviates problems such as gradient disappearance.Finally,the improved algorithm model is compared with previous representative studies,and the effectiveness of the changes such as the introduction of attention mechanism is confirmed from the subjective and objective data.(2)In order to adapt the improved image super-resolution method to the special scenario of ancient Chinese calligraphy and painting,a more complex image degradation model is designed to replace the single double triple downsampling method to obtain richer and more realistic low-resolution images at the input side;based on the purpose of processing more complex input data,the original discriminator model is replaced and the U-Net architecture is used to introduce a more fine-grained Based on the purpose of processing more complex input data,the original discriminator model is replaced and the U-Net architecture is used to introduce a finer-grained attention loss function to meticulously judge the proximity between the generated image and the actual image pixel by pixel,thus improving the performance of the generative adversarial network in general,and the effective improvement of the reconstruction results in terms of perceptual quality is experimentally verified.(3)Focusing on the practical application of the image superresolution method to the images of cultural relics painting and calligraphy,we make our own high-definition ancient painting and calligraphy image dataset,fuse the public DIV2K,Flickr2K dataset to train the improved algorithm model,and finally,through objective data comparison and visualization of the reconstructed images,we confirm the real improvement of the quality of the low-resolution painting and calligraphy images and the deeper aesthetics and artistic appreciation.