


ICDAR 2023 (International Conference on Document Analysis and Recognition), as one of the most important international conferences in the field of document image analysis and recognition, has recently received exciting news:
iFlytek Research Institute of HKUST and the University of Science and Technology of China The National Engineering Research Center for Speech and Language Information Processing (hereinafter referred to as the Research Center) won four championships in three competitions: multi-line formula recognition, document information location and extraction, and structured text information extraction.
MLHMETop: Focus on "multi-line writing", breakthrough in complexity
MLHME (multi-line formula recognition competition) test input After an image containing a handwritten mathematical formula, the algorithm outputs the accuracy of the corresponding LaTex string. It is worth mentioning that compared with previous mathematical formula recognition competitions, this competition has set "multi-line writing" as the main challenge for the first time in the industry. And unlike the previous formulas for recognizing scanned and online handwriting, this time it will recognize photographed handwriting. Mainly multi-line formulas.
The image and text recognition team of iFlytek Research Institute won the championship with a score of 67.9%, and far exceeded other participating teams in the main evaluation indicator-formula recall rate
The recall rate of the formula corresponds to the Submit Results in the list
Multi-line formulas are more complex than single-line structures, and the same character appears more often in the formula The size will also change when it appears. At the same time, the data set used in the competition comes from real scenes, and the handwritten formula pictures taken have problems such as low quality, background interference, text interference, smearing and annotation interference. These factors make the game more difficult.
The formula structure is complex and takes up multiple lines
The picture quality is not high and correction interference
Aiming at the problem that the formula has a complex structure and occupies multiple lines, the team uses Conv2former with a large convolution kernel as the encoder structure, which expands the field of view of the model and better captures the structural characteristics of multi-line formulas; innovatively proposes structured sequence decoding based on transformer The SSD processor explicitly models the hierarchical relationships within multi-line formulas in a refined manner, which greatly improves the generalization of complex structures and better models structured semantics.
In response to the character ambiguity problem caused by picture quality problems, the team innovatively proposed a semantic-enhanced decoder training algorithm. Through joint training of semantics and vision, the decoder has intrinsic domain knowledge. When characters are difficult to identify, the model can adaptively use domain knowledge to make inferences and give the most reasonable recognition results.
In response to the problem of large changes in character size, the team proposed an adaptive character scale estimation algorithm and a multi-scale fusion decoding strategy, which greatly improved the model's robustness to changes in character size.
DocILEThe crown: "Pick one in the line", the two-track competition of document information positioning and extraction topped the list
DocILE (Document Information Location and Extraction Competition) evaluates the performance of machine learning methods in locating, extracting and identifying line items of key information in semi-structured business documents.
The competition is divided into two track tasks: KILE and LIR. The KILE task needs to locate the key information location of predefined categories in the document. On this basis, the LIR task further groups each key information into different line items (Line Item), such as a single object (quantity, price) in a row in the table. )wait. iFlytek and the Research Center finally won the championship in two tracks
KILE track ranking
LIR track ranking
The left is the KILE track description, the right is the LIR track description
It can be seen from the task icon given by the official event that the information to be extracted from the document The variety is very diverse. Among them, the KILE task not only needs to extract key information of predefined categories, but also obtain the specific location of the key information; in the LIR task, a line item may have multiple lines of text in a single table. In addition, there are many types of information in the data set of this competition and the document formats are complex and diverse, which greatly increases the challenge.
The joint team proposed two technological innovation plans at the algorithm level:
In the pre-training stage, we designed a document filter based on OCR quality by extracting 2.74 million pages of document images from unannotated documents provided by the organizer. Then, we use a pre-trained language model to obtain the semantic representation of each text line in the document, and use the masked sentence representation recovery task, pre-trained under different Top-K configurations (attention span of the document in the GraphDoc model A hyperparameter)
#In the data set fine-tuning stage, the team used the pre-trained GraphDoc to extract the multi-modal representation of the text box and perform classification operations. On the basis of the classification results, multi-modal representations are sent to the low-level attention fusion module for instance aggregation. Based on the instance aggregation, the high-level attention fusion module is used to realize the aggregation of line item instances. The proposed attention fusion The modules have the same structure but do not share parameters with each other. They can be used for both KILE and LIR tasks with good results.
SVRDThe crown: No. 1 in zero-sample ticket structured information extraction task, pre-training model test
SVRD (Structured Text Information Extraction) competition is divided into 4 track sub-tasks. iFlytek and the research center are competing in the very difficult zero-sample structured information extraction sub-track (Task3: E2E zero-sample structured text extraction). Won first place
Ranking order
In the context of the official designation of key elements that need to be extracted for different types of invoices, this track requires participating teams to use model output The corresponding content of these key elements in the picture, "zero sample" means that the invoice types of the training set and the test set have no overlap; the end-to-end prediction accuracy of the track test model, the weighted average of score1 and score2 is used as the final evaluation index .
For pre-training models, zero samples put forward higher requirements. At the same time, there are various invoice formats used in the competition, and the names of the ride stops, departure times and other elements in each format are different. In addition, invoice photos have problems such as background interference, reflection, and text overlap, which further increases the difficulty of identification and extraction.
Invoices in different formats
The invoice is interfered by the striped background
The team initially adopted a copy-generate dual-branch decoding strategy to carry out the feature extraction model. When the confidence of the front-end OCR result is high, copy it directly OCR results; when the confidence level of the OCR results is low, new prediction results are generated to alleviate the recognition errors introduced by the front-end OCR model
In addition, the team also extracts sentence-level graphdoc features based on the OCR results As input to the feature extraction model, this feature integrates multi-modal features of image, text, location, and layout, and has a stronger feature representation than single-modal plain text input.
On this basis, the team also combined multiple element extraction models such as UniLM, LiLT, and DocPrompt to further improve the final element extraction effect and enable it to show better performance in different scenarios and languages. Advantages
Education, finance, medical care, etc. have been implemented to help large models improve multi-modal capabilities
This time I chose relevant events of ICDAR 2023 for the challenge, from iFlytek's real-life scenario needs in actual business; event-related technologies have also penetrated into education, finance, medical care, justice, intelligent hardware and other fields, empowering multiple businesses and products.
In the field of education, the technical ability of handwritten formula recognition is frequently used, and machines can provide accurate identification, judgment and correction. For example, the personalized precision learning and AI diagnosis in iFlytek's AI learning machine; the "iFlytek Smart Window" teaching large screen used by teachers in class, and students' personalized learning manuals, etc., have all achieved great results;
Not long ago, the Spark Scientific Research Assistant was released on the iFlytek Global 1024 Developer Festival main forum. One of the three core functions of paper reading can realize intelligent interpretation of papers and quickly answer related questions. . Subsequently, on the basis of high-precision formula recognition, the effect of organic chemical structural formulas, graphics, icons, flow charts, tables and other structured scene recognition will be advanced. This function will also better help scientific researchers improve efficiency;
Document information positioning and extraction technology is widely used in the financial field, such as contract element extraction and review, bank bill element extraction, marketing content consumer protection review and other scenarios. These technologies can realize functions such as data analysis, information extraction and comparison review of documents or files, and help business data to be quickly entered, extracted and compared, thereby improving the efficiency of the review process and reducing costs
The personal AI health assistant released on this 1024 main forum is iFlytek Xiaoyi. It can not only scan checklists and test orders and give analysis and suggestions, but it can also scan pill boxes and make further inquiries and provide auxiliary medication suggestions. For physical examination reports, users can take photos and upload them, and iFlytek Xiaoyi can identify key information, comprehensively interpret abnormal indicators, proactively inquire and provide more help. This function relies on the support of document information positioning and extraction technology
iFlytek's image and text recognition technology continues to make breakthroughs in algorithms, from single word recognition to text line recognition. to more complex two-dimensional structure recognition and chapter-level recognition. More powerful image and text recognition technology can improve the effect and potential of multi-modal large models in image description, image question and answer, image recognition creation, document understanding and processing, etc.
At the same time, image and text recognition technology is also Combining speech recognition, speech synthesis, machine translation and other technologies to form a systematic innovation, the empowered product shows more powerful functions and more obvious value advantages after application. The related project also won the first prize of the 2022 Wu Wenjun Artificial Intelligence Technology Progress Award . In the new journey, "more blooms" in several ICDAR 2023 competitions are not only feedback from iFlytek's continuous progress in the depth of image and text recognition and understanding technology, but also affirmation of its continuous expansion in breadth.
The above is the detailed content of iFLYTEK ICDAR 2023: Image and text recognition achieves greater glory again, winning four championships. For more information, please follow other related articles on the PHP Chinese website!

你可能听过以下犀利的观点:1.跟着NVIDIA的技术路线,可能永远也追不上NVIDIA的脚步。2.DSA或许有机会追赶上NVIDIA,但目前的状况是DSA濒临消亡,看不到任何希望另一方面,我们都知道现在大模型正处于风口位置,业界很多人想做大模型芯片,也有很多人想投大模型芯片。但是,大模型芯片的设计关键在哪,大带宽大内存的重要性好像大家都知道,但做出来的芯片跟NVIDIA相比,又有何不同?带着问题,本文尝试给大家一点启发。纯粹以观点为主的文章往往显得形式主义,我们可以通过一个架构的例子来说明Sam

2021年9月25日,阿里云发布了开源项目通义千问140亿参数模型Qwen-14B以及其对话模型Qwen-14B-Chat,并且可以免费商用。Qwen-14B在多个权威评测中表现出色,超过了同等规模的模型,甚至有些指标接近Llama2-70B。此前,阿里云还开源了70亿参数模型Qwen-7B,仅一个多月的时间下载量就突破了100万,成为开源社区的热门项目Qwen-14B是一款支持多种语言的高性能开源模型,相比同类模型使用了更多的高质量数据,整体训练数据超过3万亿Token,使得模型具备更强大的推

在法国巴黎举行了国际计算机视觉大会ICCV(InternationalConferenceonComputerVision)本周开幕作为全球计算机视觉领域顶级的学术会议,ICCV每两年召开一次。ICCV的热度一直以来都与CVPR不相上下,屡创新高在今天的开幕式上,ICCV官方公布了今年的论文数据:本届ICCV共有8068篇投稿,其中有2160篇被接收,录用率为26.8%,略高于上一届ICCV2021的录用率25.9%在论文主题方面,官方也公布了相关数据:多视角和传感器的3D技术热度最高在今天的开

随着智慧司法的兴起,智能化方法驱动的智能法律系统有望惠及不同群体。例如,为法律专业人员减轻文书工作,为普通民众提供法律咨询服务,为法学学生提供学习和考试辅导。由于法律知识的独特性和司法任务的多样性,此前的智慧司法研究方面主要着眼于为特定任务设计自动化算法,难以满足对司法领域提供支撑性服务的需求,离应用落地有不小的距离。而大型语言模型(LLMs)在不同的传统任务上展示出强大的能力,为智能法律系统的进一步发展带来希望。近日,复旦大学数据智能与社会计算实验室(FudanDISC)发布大语言模型驱动的中

8月31日,文心一言首次向全社会全面开放。用户可以在应用商店下载“文心一言APP”或登录“文心一言官网”(https://yiyan.baidu.com)进行体验据报道,百度计划推出一系列经过全新重构的AI原生应用,以便让用户充分体验生成式AI的理解、生成、逻辑和记忆等四大核心能力今年3月16日,文心一言开启邀测。作为全球大厂中首个发布的生成式AI产品,文心一言的基础模型文心大模型早在2019年就在国内率先发布,近期升级的文心大模型3.5也持续在十余个国内外权威测评中位居第一。李彦宏表示,当文心

不得不说,Llama2的「二创」项目越来越硬核、有趣了。自Meta发布开源大模型Llama2以来,围绕着该模型的「二创」项目便多了起来。此前7月,特斯拉前AI总监、重回OpenAI的AndrejKarpathy利用周末时间,做了一个关于Llama2的有趣项目llama2.c,让用户在PyTorch中训练一个babyLlama2模型,然后使用近500行纯C、无任何依赖性的文件进行推理。今天,在Karpathyllama2.c项目的基础上,又有开发者创建了一个启动Llama2的演示操作系统,以及一个

保险行业对于社会民生和国民经济的重要性不言而喻。作为风险管理工具,保险为人民群众提供保障和福利,推动经济的稳定和可持续发展。在新的时代背景下,保险行业面临着新的机遇和挑战,需要不断创新和转型,以适应社会需求的变化和经济结构的调整近年来,中国的保险科技蓬勃发展。通过创新的商业模式和先进的技术手段,积极推动保险行业实现数字化和智能化转型。保险科技的目标是提升保险服务的便利性、个性化和智能化水平,以前所未有的速度改变传统保险业的面貌。这一发展趋势为保险行业注入了新的活力,使保险产品更贴近人民群众的实际

杭州第19届亚运会不仅是国际顶级体育盛会,更是一场精彩绝伦的中国科技盛宴。本届亚运会中,快手StreamLake与杭州电信深度合作,联合打造智慧观赛新体验,在击剑赛事的转播中,全面应用了快手StreamLake六自由度技术,其中“子弹时间”也是首次应用于击剑项目国际顶级赛事。中国电信杭州分公司智能亚运专班组长芮杰表示,依托快手StreamLake自研的4K3D虚拟运镜视频技术和中国电信5G/全光网,通过赛场内部署的4K专业摄像机阵列实时采集的高清竞赛视频,


Hot AI Tools

Undresser.AI Undress
AI-powered app for creating realistic nude photos

AI Clothes Remover
Online AI tool for removing clothes from photos.

Undress AI Tool
Undress images for free

Clothoff.io
AI clothes remover

AI Hentai Generator
Generate AI Hentai for free.

Hot Article

Hot Tools

Notepad++7.3.1
Easy-to-use and free code editor

Atom editor mac version download
The most popular open source editor

Dreamweaver Mac version
Visual web development tools

Dreamweaver CS6
Visual web development tools

DVWA
Damn Vulnerable Web App (DVWA) is a PHP/MySQL web application that is very vulnerable. Its main goals are to be an aid for security professionals to test their skills and tools in a legal environment, to help web developers better understand the process of securing web applications, and to help teachers/students teach/learn in a classroom environment Web application security. The goal of DVWA is to practice some of the most common web vulnerabilities through a simple and straightforward interface, with varying degrees of difficulty. Please note that this software
