AI-based document preprocessing for optical character recognition

尽管自动光学字符识别（OCR）技术已经成熟，但最近面临着新的挑战：文件来源现在来自包括移动相机在内的多个来源拍摄的照片。这些图像包含光照、失真和其他降低OCR算法精度的人为因素的显著变化。在本研究中，将OCR建模为一个未知盒，并采用适合于未知盒的基于深度学习的算法来提高图像质量。未知框假设准确地反映了最终用户无法访问OCR组件的现实，例如，专有软件。我们提出的提高图像质量的方法依赖于图像分析和计算机视觉算法的最新进展。这项研究的行业合作伙伴是Inuit， Inc.和Intuit，加拿大，他们是自动化文档处理的领导者之一。

Even though automated optical character recognition (OCR) is now a mature technology, it is facing a new challenge these days: sources of documents are now coming from pictures taken by multiple sources including mobile cameras. These images contain significant variations in lighting, distortion, and other artifacts that lowers the accuracy of OCR algorithms. In this research, model OCR as an unknown-box and improve image quality by deep learning-based algorithms suitable for the unknown-box. The unknown-box assumption accurately reflects a reality where end-users do not have access to the components of the OCR, e.g., proprietary software. Our proposed method to improve image quality relies on state-of-the-art advancements in image analysis and computer vision algorithms. The industry partner for the research is Inuit, Inc. and Intuit, Canada who are one of the leaders in automated document processing.