| Paper documents are also one of the main ways of transmitting information in the Internet era,but there are many disadvantages to using paper documents to transmit information,such as inconvenient transmission,easy loss and inconvenient preservation.With the continuous development of OCR(Optical Character Recognition)technology,the conversion of paper documents into electronic information is becoming more and more accurate and is being used in more and more scenarios in life.The scanning pen derived from OCR technology is mainly responsible for the acquisition of text images and the display of OCR recognition results.In view of the limited pixels of text images acquired by the scanning pen image,the limited window size and the poor quality of the images acquired by handheld sliding,how to stitch the fragmented text images into a complete,text-informed,high-quality image,as well as to meet people’s daily life How to stitch together fragmented text images into a complete,high quality image containing text information,and to meet the real-time requirements of people’s daily lives,is a problem that needs to be solved by current scanning pen products.This paper focuses on the analysis of continuous multi-frame text images and improves on the traditional template matching algorithm by combining it with other image processing algorithms.The paper’s research is divided into the following main areas:(1)In response to the problems that the traditional template matching method uses the previous stitching result as the next input,resulting in the need for a large number of pixels to be repeatedly copied,the template scanning range is too large,and there are invalid images in the images,this paper adopts three methods to filter the invalid images,such as image entropy,edge information and variance,etc.,takes the relative alignment points of two adjacent frames,and adds a y-axis offset to the search image to make the image alignment more The image alignment is made more accurate by adding a y-axis offset to the search image.The stitching result is detected as feature points,a straight line is fitted to the feature points to obtain the inclination angle,and the image is rotated for correction.(2)To address the problem that the traditional template matching algorithm requires global scanning,the character concatenated domains are used as templates to form concatenated domain chains by pairing adjacent images.As there may be more than one connected domain for a single character,this paper merges the connected domains and uses least squares estimation to aggregate the connected domains into a text line,iteratively updates the squares estimation parameters,and finally completes the skew correction of the text image by translation.In order to address the problem of multiple matches in the connected domain matching,this paper uses a projection alignment algorithm for local matching correction,and finally selects the best connected domain in each connected domain chain as the final stitching result.The improved scheme proposed in this paper has been experimentally validated.In terms of OCR recognition result string similarity,the improved scheme based on the concatenated domain has increased the cosine similarity by 0.0214,Jaccard correlation coefficient by 0.016,edit distance by 0.8681,Needleman-Wunsch algorithm score by1.4489 and Smith-Waterman score by 0.707 compared to the improved scheme based on the traditional template matching algorithm.The image stitching was able to meet the real-time stitching requirements in terms of time. |
PDF Full Text Request