Open set text recognition technology

开放环境下的模式识别与文字识别应用中，新数据、新模式和新类别不断涌现，要求算法具备应对新类别模式的能力。针对这一问题，研究者们开始聚焦开放集文字识别（open-set text recognition，OSTR）任务。该任务要求，算法在测试（推断）阶段，既能识别训练集见过的文字类别，还能够识别、拒识或发现训练集未见过的新文字。开放集文字识别逐步成为文字识别领域的研究热点之一。本文首先对开放集模式识别技术进行简要总结，然后重点介绍开放集文字识别的研究背景、任务定义、基本概念、研究重点和技术难点。同时，针对开放集文字识别三大问题（未知样本发现、新类别识别和上下文信息偏差），从方法的模型结构、特点优势和应用场景的角度对相关工作进行了综述。最后，对开放集文字识别技术的发展趋势和研究方向进行了分析展望。;Text recognition is focused on transcription-based image processing modeling in relevance to such domains like document digitization, content moderation, scene translation, automation driving, understanding, and other related contexts. Conventional techniques are often concerned about characters-seen more. However, two factors the training set of these methods yet be well covered, which novel character categories out-of-vocabulary (OOV) samples. Newly characters-related samples linked with OOV-based it may pay attention seen characters without combinations or For categories, internet-based environments can mainly used face unseen ligatures 1) emoticons unperceived languages, 2) scene-text environments, 3) from foreign region-specific languages. digitization profiling, undiscovered not involved as well. Since heterogeneity language format balanced, linguistic statistic data(e. g., n-gram, context, etc.) biased data gradually, challenged for vocabulary-high-correlated methods. The required yield three key scientific problems that affect costs efficiency open-world applications. oriented spotting capability, whereas characters-unseen rejected replace silent characters. Furthermore, popular open-set problem, leaked out mentioned below. First, emergence efficient many cases, re-training upon each occurrence costly, an incremental learning capability need strengthened after that. Second, amount received generalized zeroshot task. Third, Linguistic bias robustness yielded by OOV Due characterbased nature prediction, more possess handle some extent. capabilities constrained demonstrate strong vocabulary reliance because capacity models, recognition(OSTR) task feasible since existing tasks zero-shot model individual aspects only. This aims spot recognize characters, robust skews. As extension conventional task, OSTR retain a decent contents. In recent years, has been developing intensively context recognition. literature review carried its domains. It consists five background, genericity, concept, implementation, summary. we introduce application background analyze specific OSTR-derived cases. generic introduced brief preliminary less familiar researchers field. definition introduced, followed discussion relationship tasks, e. close-set Its implementation-wise, common frameworks first introduced. recognized derivations frameworks, where derivation based following:new category spotting, classes, robustness. Specifically, new problem refers rejecting come absent class given label set. Slightly different label-set straightfoward. Incremental terms non-retrained side information corresponding categories. slightly definition, excluded generative adversarial network(GAN)-based transductive approaches. holds original beyond stressed solution covered modeling-similar fields. evaluation cover datasets protocols contexts listed:1) multiple public available datasets, commonly metric measure performance, several protocols, typical methods, performance. Here, protocol compositions sets, testing metrics. summary, comparative analysis growth technical preference demonstrated. Finally, potnetialss trends future research directions predicted further.