| In computer vision technology,digital image processing has become a hot research field.Among them,super-resolution technology aims to establish mathematical models from low-quality blurred images and use this model to fit degraded images to reconstruct clearer and more detailed image information.This technology is of great significance for the research of image processing and optical character recognition.Due to various factors such as environmental factors and transmission media,the quality of some precious ancient book images has been compromised.This paper studies and explores the problems in super-resolution reconstruction of the historical Tibetan document images.The main content is as follows.(1)due to the lack of the historical Tibetan document images datasets for superresolution technology research,a complex degradation model was used to construct the image pair dataset in the paper,laying a data foundation for super-resolution research;A Tibetan character recognition dataset was constructed using text line segmentation method to verify the effectiveness of super-resolution technology in improving the accuracy of text recognition.(2)based on the Zero Shot SR(ZSSR)self-supervised model,an improved EZSSR network model is proposed to achieve super-resolution reconstruction of multi-scale historical Tibetan document images.This method can quickly achieve super-resolution reconstruction of a single image,using the feature pyramid structure in the Laplace Pyramid Super Resolution Network(Lap SRN)to reconstruct super-resolution images of different sizes.The model has two branch networks: feature extraction branch and reconstruction branch,which jointly learn residual features and perform up-sampling reconstruction,accelerating the convergence speed of the network.(3)an improved ESRGAN+ model is proposed using the super-resolution method of generating adversarial networks(SRGAN)to achieve super-resolution reconstruction of the historical Tibetan document images that are more in line with human perception.This method utilizes a generative adversarial network structure,with an objective function using perceptual loss.The model structure is divided into a generator and a discriminator.The generator consists of two parts: one is a feature extraction and up-sampling layer using a dense residual block RRDB structure,and the other is an image enhancement layer using a weighted least squares filter WLS.The discriminator adopts a more stable U-Net model with spectral normalization.(4)various image quality evaluation methods have been adopted for highresolution images reconstructed by super-resolution algorithms,including subjective evaluation method: average subjective opinion score MOS,objective evaluation method: peak signal-to-noise ratio PSNR,structural similarity SSIM,and natural image quality evaluation NIQE.The quality of image reconstruction cannot be comprehensively and accurately evaluated by just one evaluation method,so multiple evaluation methods have been adopted to comprehensively consider image quality to make it more reasonable.(5)a convolutional recurrent neural network text recognition method based on variable scale mechanism was used to conduct end-to-end Tibetan character recognition research on images reconstructed using different super-resolution algorithms.The recognition results effectively reflect the advantages and disadvantages of superresolution algorithms. |
