Digital access and transcription of historical and complex document images

根据联合国教育、科学及文化组织(教科文组织)的说法，遗产是我们过去留下的遗产，是我们今天生活的东西，也是我们留给下一代的东西。自1972年联合国教科文组织通过一项国际条约以来，鼓励识别、保护和保存世界各地优秀遗产的持续努力一直在进行中。特别是那些关注视觉文化遗产的人，如古代文献，从信号处理(SP)和成像技术的进步中受益匪浅。使这些文件清晰易懂是一项艰巨但重要的任务，特别是在它们具有重大历史意义的情况下。这些文档可能会遭受多种形式的恶化和物理退化现象(背景噪声、老化造成的伪影、干扰图案以及笔划的渗漏或阴影)，它们可能被故意擦除和覆盖，并且可能存在重叠。拟议研究方案的目标是确定一种实现系统主要组成部分的方法，该系统旨在增强古代文献的内容并恢复用于检索这些文献的数字图像的隐藏信息。拟议的方法将制定一个统一的图像分析和分类框架。为了实现高质量的复原，将发展包含随机过程的真实物理图像退化模型来模拟退化现象。然后，我们将能够设计出有效的反向扩散模型，能够直接恢复这些类型的物理扩散。为了达到较高的搜索能力，我们将使用增量学习，将专家(学者和历史学家)放在循环中对子空间进行数据分类和统计主题类分类，这将减少搜索空间，提高系统效率。我们的长期目标是建立可扩展的智能系统，具有类似人类的能力，可以无缝访问和利用现有的大量和高度退化的多语言文档。

According to the United Nations Educational, Scientific and Cultural Organization (UNESCO), "Heritage is our legacy from the past, what we live with today, and what we pass on to the future generation." Since an international treaty was adopted by UNESCO in 1972, an ongoing effort to encourage the identification, protection, and preservation of outstanding heritage around the world has been under way. Particularly those focusing on visual cultural heritage, like ancient documents, have greatly benefited from signal processing (SP) and imaging advances. The task of making such documents legible and accessible is a daunting, but important one, especially when they have great historical significance. These documents may suffer from many forms of deterioration and physical degradation phenomena (background noise, artifacts caused by aging, interfering patterns, and the bleed-through or shadow-through of strokes), they may have been deliberately erased and overwritten, and there may be overlays. The objective of the proposed research program is to define a methodology for the realization of the principal components of a system designed to enhance the content of ancient documents and recover hidden information for the retrieval of digital images of those documents. The proposed methodology will devise a unified image analysis and classification framework. To achieve high-quality restoration, true physical image degradation modeling, which includes stochastic processes, will be developed to model the degradation phenomena. We will then be able to devise efficient reverse diffusion models capable of directly restoring these types of physical diffusion. To achieve high search capability, we will use incremental learning, putting experts (scholars and historians) in the loop to perform data classification and statistical topical class categorization in subspaces, which will reduce the search space and increase system efficiency. Our long-term goal is to build scalable intelligent systems with the human-like ability to seamlessly access and make use of the huge volume and variety of ancient and highly degraded multilingual documents that exists.