Software Evaluation Method of Optical Character Recognition System Based on Metamorphic Testing

光学字符识别系统通过扫描图像，使得原本只能以图像形式存在的文字信息可以被计算机处理和分析，已在各领域广泛应用。光学字符识别技术的质量是重点，如何系统地对其质量进行评估仍是一项具有挑战性的工作，因为它的评估方法依赖于大量的标记数据，同时也依赖于人工判断。在本文中，从用户的角度对光学字符识别系统进行评估，帮助用户更好地了解这些系统的功能，进而选择适合自己的光学字符识别系统，以满足他们的特定需求。为了更好地评估光学字符识别系统，光学字符识别系统的特征，从用户角度出发，我们定义了两类蜕变关系模式。在此基础上，定义了六个抽象的蜕变关系，并选取了光学字符识别的广泛应用场景车牌识别来评估系统的质量。为保证在没有先验标注的情况下评估OCR系统的性能以及鲁棒性，本文提出了此方法。Optical Character Recognition (OCR) systems enable computers to process and analyze text information that originally exists only in image form through scanning, and have been widely adopted across various domains. The quality of OCR technology is of paramount importance, yet systematically evaluating its performance remains challenging as current assessment methods heavily rely on large volumes of labeled data and manual judgment. In this paper, we propose evaluating OCR systems from the user perspective to help users better understand system capabilities and select appropriate OCR solutions that meet their specific needs. To facilitate comprehensive evaluation, we define two categories of metamorphic relation patterns based on key OCR characteristics from the user perspective. Furthermore, we establish six abstract metamorphic relations and validate them through license plate recognition—a prevalent application scenario for OCR technology. The proposed method enables performance and robustness evaluation of OCR systems without requiring prior annotated data.