Research And Application Of Text Image Super-resolution Based On SRGAN

Image super-resolution is the process of using low-resolution images to generate high-resolution images of the same scene.Traditional super-resolution mostly uses interpolation algorithm,which has limited improvement on the quality of reconstructed image.In recent,image super-resolution based on deep learning framework is the focus of current research and has achieved better results.However,when the super-resolution algorithm based on the deep learning framework is applied to the text image,the reconstructed image has the problems of poor edge recognizability due to blurred edges and handwriting adhesion.Aiming at this problem,this paper proposes an edge feature fusion text image super-resolution algorithm based on the characteristics of text images,and analyzes and studies the implementation and application of the algorithm.(1)Aiming at the problem that the text edge is suppressed during the edge extraction process,a large amount of text edge information in the reconstructed image is lost.In this paper,a regional blur strategy based on gray threshold is proposed.Based on the gray level in the neighborhood of the central pixel,the regional blur threshold and fuzzy weight are obtained,and the blurred gray value is calculated.This method can not only keep more font edges,but also can automatically adjust the fuzzy weight according to the gray level of the area.This paper compares this algorithm with the edge extraction algorithm based on deep learning framework.Under the premise of ensuring similar edge extraction effects,the calculation time is almost 2‰ of the latter.(2)The low-resolution text image contains fewer effective features,resulting in the lack of detail in the reconstructed image and the blurred edges of the text in the reconstructed image.This paper improves the feature extraction module in the SRGAN network and introduces a shared source residual group structure.Use global operations to capture remote spatial context information,and use multiple local residual attention groups to learn more abstract features.The edge feature fusion and the edge enhancement loss function based on the edge gradient weight are used to enhance the edge quality of the reconstructed image.The algorithm achieves the effect of enhancing the edge by appropriately amplifying the loss of the reconstructed image and the reference image at the edge position.In addition,the text introduces a second-order attention network for more powerful feature expression and feature relationship learning.Through a trainable second-order channel module,the feature channel is adaptively adjusted to enhance the influence of edge features on the reconstructed image.In this paper,multiple experiments are designed on ICDAR-2015 and other data sets to verify that the use of edge feature fusion,edge enhancement loss function and the introduction of shared source residual group with second-order channel attention mechanism are better for improving the quality of reconstructed text images promotion.In addition,the text image super-resolution reconstruction platform is implemented according to the algorithm design in this paper,which provides online super-resolution reconstruction service of text images.