Optical Character Recognition (OCR) has a significant application in obtaining text from academic video material, particularly from lecture slides. Still, most of the available OCR assessments address documents holistically and do not consider structural and semantic variance contained in slide content. This paper comprehensively benchmarks five open-source OCR engines—Tesseract, EasyOCR, PaddleOCR, Keras-OCR, and DocTR—on labeled semantic regions of lecture slides, including titles, text boxes, tables, handwritten notes, headers, and footers. Because of architecture and runtime constraints, DocTR and Keras-OCR were excluded from the final performance comparison. The study examines OCR engine performance over these region categories using Word Error Rate (WER) and Character Error Rate (CER) as metrics. Findings indicate no one OCR engine stands out across categories: Tesseract works consistently on formatted text areas such as titles and headings, while PaddleOCR is best at identifying handwritten and tabular data. The results emphasize the necessity of region-aware OCR selection in applications for indexing lecture videos. This contribution offers a pragmatic benchmark and actionable recommendations for researchers and engineers constructing searchable educational content platforms.
Purushotham et al. (Wed,) studied this question.