SBIR Phase I: Real-time, accurate OCR from Video using Intra- and Inter-Frame Machine Learning

该小型企业创新研究(SBIR)第一阶段研究项目专注于开发开创性的实时算法，用于使用字体和文本字符串的机器学习在复杂3-D环境的数字视频中自动查找和识别文本。从本质上讲，该项目将OCR从一项用于2-D文档的技术带到了3-D世界。该项目建立在文档的光学字符识别(OCR)算法的基础上，传统的OCR在这些地方无法实现：彩色小册子、杂志封面和其他来源，其中照片、线条艺术和任意旋转的文本使OCR过程变得非常复杂。该项目旨在建立在这项技术的基础上，找到在复杂的3D现实世界场景中查找和识别文本的解决方案，例如路牌和店面，在这些场景中，文本可能与相机的任何3D角度成任意角度。这个项目成功的关键是该算法对字体的机器学习能力。有许多令人兴奋的应用程序受到来自视频源的准确OCR的影响。虽然可以对视频源中的文本进行OCR，但它通常必须在明显的文本上进行，例如字幕，并且不能实时完成。实时和准确的视频OCR将支持以下应用：1)通过其中包含的文本对数字视频片段进行自动索引，2)帮助盲人在室内和室外独立导航，3)自动连续路边或基于车辆的车牌扫描，以及4)作为地面真实信息，以提高GPS的精度。因此，这项技术的市场包括个人、公司和政府机构。其社会影响包括1)使数字化视频库能够以更低的成本通过更多的元数据标签进行搜索，2)为盲人提供更大的独立性和安全性，3)通过自动识别被报告被盗的汽车或被吊销执照的人拥有的汽车来改善道路安全，以及4)提高GPS导航的准确性。技术影响将出现在机器学习应用于视频OCR、实时OCR和低分辨率OCR的领域。

This Small Business Innovation Research (SBIR) Phase I research project focusses on the development of ground-breaking real-time algorithms for automatically finding and recognizing text in digital video of complex 3-D environments using machine learning of fonts and text strings. Essentially, the project takes OCR from being a technology for 2-D documents and brings it to the 3-D world. The project builds on algorithms for optical character recognition (OCR) of documents where conventional OCR fails: colorful brochures, magazine covers, and other sources where photographs, line art, and arbitrarily-rotated text greatly complicate the OCR process. The project aims to build on this technology to find solutions to the finding and recognizing text in complex 3-D real world scenes such as street signs and storefronts where the text may be at any arbitrary 3-D angle to the camera. Critical to the success of this project is the algorithm's capability for machine learning of fonts. There are a number of exciting applications that are impacted by accurate OCR from video sources. While OCR of text in video sources can be done, it usually must be on plainly obvious text, such as subtitles, and it cannot be done in real-time. Real-time and accurate video OCR would enable applications that include 1) Unaided indexing of digital video footage by the text contained therein, 2) aiding the blind navigate independently, both indoors and outdoors, 3) automated continuous roadside or vehicle based license plate scanning, and 4) as ground truth for improved GPS accuracy. Markets for the technology therefore include individuals, corporations, and government agencies. The societal impacts include 1) rendering digitized video libraries searchable by more metadata tags at low cost, 2) greater independence and safety for the blind, 3) improving road safety through automatically identifying cars reported stolen or cars owned by people with suspended licenses, and 4) improved GPS navigation accuracy. Technological impacts will be in the areas of machine learning applied to video OCR, real-time OCR, and low-resolution OCR.