Research On Digital Art Image Super-Resolution Reconstruction Method Based On Deep Learning

Thanks to the development of blockchain technology,digital art images have gained new vitality.Digital art images are bought,sold,and traded on art platforms built on blockchain,with ownership of the artworks recorded on the blockchain.The new generation of digital art images made possible by blockchain technology have limited supply and unique scarcity.As the internet continues to develop,more and more artists and collectors are experimenting with new art forms,and digital art images are becoming increasingly popular.There are several phenomena in the field of digital art images: First,a significant portion of digital art images are pixel-style artworks.The challenge is to increase the clarity and display quality of these images without breaking the original detail and artistic style.Second,many existing digital art images have low pixel counts.Due to the unique characteristics of art,the appearance of these works cannot be easily altered.When collectors want to print and frame their collections for display at home,they face the problem of image blurring due to low pixel counts.To address these phenomena and challenges in the field of digital art images,this paper uses the SRGAN network model as a benchmark and made the following contributions to improve the super-resolution of digital art images:(1)This paper refers to current popular art styles and collects a large number of Non-Fungible Digital Assets(NFDAs)licensed under the Creative Commons Zero(CC0)agreement to propose a digital art image dataset.Digital art images have unique textures and patterns,and a distinct dataset helps to maintain artistic style while obtaining clearer super-resolution images,thus increasing their value and influence.(2)This paper proposes the High-Frequency Residual Attention SuperResolution Generative Adversarial Network(HFRA-SRGAN)model,which optimizes the generator’s network structure.The Res UNet replaces some residual modules in the generator,enabling the generator network to fuse multi-scale features of the image,enhancing its feature extraction capability and obtaining more image details.To achieve better super-resolution performance,the batch normalization layer in Res Net50 is removed,reducing unnecessary parameters,speeding up model training,and enhancing generalization.(3)This paper introduces the High-Frequency Residual Attention(HFRA)module,used in the generator’s Res UNet.By combining the HFRA module with Res UNet and leveraging skip connections,the generator can integrate multi-scale feature information and high-frequency information,supplementing the model with richer details,enhancing image super-resolution quality,and improving the reconstruction effect.(4)Based on the HFRA module,this paper introduces a high-frequency loss function.This loss function focuses on learning high-frequency details,improving the clarity and fidelity of super-resolution images.It also retains the style details of the input image,helping the model generalize the overall style,and is particularly suitable for digital art images,enhancing super-resolution reconstruction results.Experiments were conducted on the digital art image dataset with both the proposed HFRA-SRGAN model and the benchmark model.Under the same conditions,the experimental results show that the image super-resolution reconstruction performance and evaluation metrics of the HFRA-SRGAN model are superior to the benchmark model.The HFRA-SRGAN model achieves a Peak Signalto-Noise Ratio(PSNR)of 24.19 and a Structural Similarity(SSIM)of 0.6038,which are 6.7% and 1.1% better than the benchmark model,respectively.