The Research And Application Of Text-image Super-resolution Based On Regularization

It will lead to poor image quality and the blurred edge because of theimpact of imaging hardware restrictions or conditions. Unilaterally improvingthe precision of imaging hardware will not only increase the cost of the product,but also can not completely remove the interference of imaging environment.Undoubtedly, super-resolution reconstruction technique is the best way to solvethis problem, which not only can avoid the waste caused by improved imagingsystem, but also can effectively improve the image quality.So far, the super-resolution reconstruction techniques have been deeplystudied and widely used. However, super-resolution reconstruction algorithmspecifically for document image is still a minority, and there are many flaws inthe existing reconstruction algorithms of document images. For this, this articlemakes a thorough study and comprehensive analysis for the theory ofreconstruction algorithm based on regularization of document imagesuper-resolution, and the application of technology, the main contents include:(1)We have studied the difference between the natural image and textimage, and have analyzed the two unique properties of text image: the bimodaldistribution of gray values and the gradient values piecewise smooth. Then based on these characteristics, mathematical reconstruction model is presentedfor document image.(2)It reveals the nature of the process of calculating the optical flow fieldthat, it is a procedure for solving linear systems, by specific research on tieredpyramid principle and optical flow estimation method. Finally, based on theclassical optical flow calculation method, this paper has improved the old one,which has used adaptive iterative method instead of the traditional first-orderTaylor series. The experimental results show that the proposed algorithm in thisarticle gets a higher calculation accuracy than the old one, which is fullyapplicable to the motion estimation in document image reconstruction process.(3)We have analyzed the function feature of Geman&McClure norm, andhave added it to the objective function as a data fitting items, and havedemonstrated its superior robustness by experiments. Derived theoretical basisof bilateral total variation BTV regularization term, the model introduces prioriHuber function into regularization term design. Finally, compared with theexisting algorithm according to the proposed algorithm experiments, the resultshave verified that the characteristic of the proposed algorithm can effectivelyemploy low-resolution document image and use the priori knowledge ofstructural features of the character reasonably. All Explain that the algorithm canbe more effectively protect the edge in the same application environment in thispaper, remove the noises of all types, significantly enhance the character recognition rate of low-resolution document image while reducing the executivetime, applicable to all types of low-resolution document image.(4)We have analyzed several related important modules of the license platerecognition, including image block edge detection, regional location andcharacter segmentation. Then the proposed algorithm has been reasonablyintroduced in front the steps of license plate character segmentation to improvethe character image resolution. Finally, the experimental results have showed thealgorithm is applicable to the far horizon license plate recognition: it caneffectively enhance the license plate image resolution of distant horizon, andincrease the applicable distance of license plate recognition, and greatly enhancethe overall performance of the license plate recognition system.