Semi-structured transaction documents acquired by scanners and cameras exhibit different degradation patterns. This study evaluates Otsu and Sauvola as baseline binarization methods on 100 self-collected document images, balanced between scanner and camera acquisitions (50 images each). Each image was converted to grayscale, enhanced with CLAHE, processed using six limited parameter configurations, and compared with manually validated pixel-level binary ground truth. Performance was assessed using MSE, PSNR, SSIM, precision, recall, F1-score, foreground ratio, and DRD, with F1-score as the primary criterion. Otsu_T10 achieved the best scanner-domain performance (F1-score = 0.917; DRD = 0.166). For camera images, Otsu_T11 produced the highest F1-score (0.172), while Sauvola_k03 yielded the lowest DRD (0.136). Visual failure analysis showed that scanner errors were dominated by broken thin strokes or character thickening, whereas camera errors included disconnected characters, missing small digits and table lines, and thick foreground artifacts caused by shadows and dark borders. These results confirm that the tested methods are baseline approaches rather than a complete camera-document processing solution. Perspective-aware cropping or correction, illumination normalization, and shadow reduction are mandatory before thresholding and downstream OCR on camera images. The conclusions are limited to the tested dataset and parameter ranges.
Copyrights © 2026