Comparative Evaluation of Thresholding Methods for Optimized Digital Document Parsing Accuracy

Automated parsing of semi-structured documents has become increasingly important, particularly in standardized formats like SATS-LN, which contain fixed-layout fields such as permit number, addresses, validity period, and item types. This study investigates the impact of two thresholding methods Otsu and Sauvola on object detection accuracy using Faster R-CNN with Detectron2. A dataset of 200 SATS-LN documents, captured via scanner and camera, was augmented into 3,600 images and labeled for seven key fields. Image quality was evaluated using PSNR, SSIM, and MSE, while detection performance was measured through mAP, AP50, AP75, AR@100, precision, recall, and F1-score. Results showed that Sauvola preserved structural layout more effectively (SSIM: 0.76 for scanner, 0.47 for camera), although Otsu achieved higher PSNR on scanned images. Sauvola attained the highest macro and weighted F1-score (0.998), with near-perfect label detection and consistent performance across augmentations. Overall, Sauvola is more reliable for enhancing segmentation and detection in layout-based document processing.