| Improper preservation of ancient Tibetan books over a long period of time and other factors will result in color degradation of the text and stains on the layout.In addition,since most of the ancient Tibetan books are in the shape of large long books,the document images are mainly obtained by taking photos.When photographing ancient Tibetan books in low illumination environment or under technical restrictions,the images will be darker.The details are hidden in the shadows.Therefore,it is very necessary to enhance the document image of ancient Tibetan books,which can not only improve readers’ reading experience,but also lay a good foundation for subsequent research work such as document image analysis and recognition.The main work is as follows:1.Image data set construction of low-illumination Tibetan ancient books.Existing low-light image enhancement works mainly focus on real scene images,and few focus on document images.Moreover,since the document images of ancient Tibetan books are not easy to obtain and there is no public low-light document image data set,the contrast adjustment strategy driven by prior knowledge and the method based on light transfer are used to construct the low-light Tibetan ancient books document image data set.2.Study on image enhancement methods of Tibetan ancient books without supervised learning.By improving the network structure and loss function,more prior knowledge reflecting the environmental characteristics is introduced to realize the effects of low-light image enhancement,background denoising and text highlighting,and improve the quality of document image.An improved unsupervised image enhancement generation Antagonism network(UHTEGAN)and a zero-reference residual attention-mechanism depth curve estimation network(Zero-RADCE)for image enhancement of low-illumination Tibetan ancient books are proposed respectively.Through experimental research,compared with the original method,UHTEGAN and Zero-RADCE have improved qualitative and quantitative evaluation indexes to some extent.Although UHTEGAN does not need to train paired data sets,high-quality image guide generator is needed to improve and enhance the effect,while Zero-RADCE only needs low-illumination images for training.Compared with UHTEGAN,the number of model parameters is reduced by nearly 90%.3.Enhance the quality assessment of document images.The quality of enhanced images of ancient books was evaluated subjtically and objectively.PSNR,SSIM and MSE in full-reference quality evaluation were used for objective evaluation,while BRISQUE in no-reference quality evaluation was used for objective evaluation.In order to further test the enhancement effect,the binarization effect of the enhanced document image and the improvement of the recognition rate in character recognition are discussed. |
