Abstract
Binarization of fiscal receipts becomes difficult when the thermal printing has faded and lighting is uneven. We investigate the classical local rule Mean , , which requires only a single integral representation or one set of line buffers. The authors analyze when its absolute shift outperforms the relative shift of the Bradley–Roth method or the variance-based shift of the Sauvola method. The image formation model defines the permissible band (noise level , ink fraction p, ink contrast ). The model explains why Mean thresholding preserves faded print and why it fails in the presence of strong noise. Therefore, the authors determine the shift by a robust noise estimate, (noise-adaptive Mean-C, NA-Mean-C), without changing the rules for an individual pixel. In the benchmark, the parameters for 11 methods were selected using a separate validation set; the results were evaluated across five degradation categories without exceptions. On 150 synthetic receipts, NA-Mean-C achieved the highest average F-measure (91.61%) and the lowest average character error rate (CER) (6.89%), whereas Mean-C with a tuned fixed shift failed to handle high-frequency noise (F-measure 15.89%). The Wolf and Gatos methods remained the best in specific categories. On 80 public Scanned Receipt Optical Character Recognition (OCR) and Information Extraction (SROIE) receipts—across fourconditions—the NA-Mean-C method and other mean-based approaches successfully extracted the total amount from 55.6 to 57.5% of the documents compared to 48.8% for Gatos and 40.3% for Otsu. A bit-accurate model of the streaming implementation shows that, compared to Sauvola, Mean saves only about 6% of on-chip memory but does not require a general-purpose multiplier, a divider, or the calculation of a square root. Evaluated independently of binarization and OCR, the keyword-based total amount extractor improved accuracy on the SROIE benchmark text from 42.9% to 83.0%.