Extracting textual information from video recordings remains challenging due to variations in the size, layout, and visual orientation of each image frame, which can lead to character recognition and word segmentation errors in Optical Character Recognition (OCR). This study aims to develop a post-processing framework for improving the quality of OCR-generated text by addressing over-segmentation and minor character recognition errors. The main contribution is the Word Segmentation Recovery Lexicon Approach (WSR-LA), which uses an Indonesian lexicon to reconstruct over-segmented words prior to character-level correction. The reconstructed text is subsequently corrected using similarity-based matching to identify the most appropriate lexicon candidates. An ablation study was conducted on 45 video frames to compare Levenshtein Distance (LD), Jaro (Jr), Jaro-Winkler (Jrw), Jaccard (Jcd), and Damerau-Levenshtein (DL) similarity methods at different correction thresholds. The LD method with a 70% threshold achieved the best performance, reducing the average Character Error Rate (CER) to 0.95% and Word Error Rate (WER) to 3.89%. The proposed WSR-LA framework was then evaluated on 25 additional video frames and achieved an average CER of 0.63% and WER of 1.96%, representing the lowest error rate among the four evaluated rule-based and lexicon-based baseline methods. These results demonstrate that the proposed framework effectively addresses minor OCR character errors, including insertions, deletions, and substitutions, as well as word over-segmentation. The framework can therefore support more reliable extraction of textual information from video frames for downstream video analysis and retrieval applications.
Copyrights © 2026