基于GPT模型语义反馈的多切分选择算法  

A Multiple Segmentation Selection Algorithm Basedon GPT Model Semantic Feedback

作　　者：许美佳 孙羽菲 王娜 张玉志 Xu Meijia;Sun Yufei;Wang Na;Zhang Yuzhi(School of Software,Nankai University,Tianjin 300450,China)

机构地区：[1]南开大学软件学院,天津300450

出　　处：《南开大学学报（自然科学版）》2023年第6期19-23,共5页Acta Scientiarum Naturalium Universitatis Nankaiensis

基　　金：国家重点研发计划(2021YFB0300104)。

摘　　要：切分是光学字符识别中一个至关重要的步骤,它直接影响着最终文本识别的准确性.传统光学字符识别一般是针对给定的文本图像提供单一的切分策略,但在处理低质量或者格式多变的图像时效果较差.研究了多切分候选的选择问题,提出了一种针对多切分候选的基于GPT模型语义反馈的切分选择算法.实验表明,本文提出的算法,通过与GPT模型的语义评估能力有效结合,对低质量文本图像的切分任务取得了显著的改进,进一步提升了识别的准确率.Segmentation is a crucial step in OCR,which directly affects the accuracy of final text recognition.Traditional OCR generally provides a single segmentation scheme for a given text image,but its effectiveness is poor when dealing with low-quality or format changing images.The selection problem of multiple segmentation candidates is studied and a segmentation selection algorithm is proposed based on GPT model semantic feedback for multiple segmentation candidates.Experiments have shown that the algorithm proposed here has significantly improved the segmentation task of low-quality text images by effectively combining it with the semantic evaluation ability of the GPT model,further enhancing the accuracy of recognition.

关 键 词：GPT模型 语义反馈 多切分 分支限界法 

分 类 号：TP311.5[自动化与计算机技术—计算机软件与理论]