A Lightweight Spatial-Anchor-Aware Confidence Routing Method for Heterogeneous Scanned Documents

Abstract

Automatic routing is a necessary first stage in many scanned-document workflows, yet small systems often face an unattractive choice between brittle keyword rules and computationally expensive deep classifiers. Whole-page rules discard word placement, while a forced prediction can silently misroute an ambiguous page. This paper presents SACR, a lightweight Spatial-Anchor-Aware Confidence Router. Class-specific anchors are induced from OCR text with a smoothed document-frequency score and discounted when their training support is weak. Anchor evidence is adjusted by normalized top, middle, and bottom page regions, then combined with ten inexpensive layout statistics. A top-two score margin sends uncertain pages to human review. Experiments use a deterministic, class-balanced, public archival subset of corrected Tobacco3482 annotations with 400 training, 125 validation, and 125 test pages from five categories. SACR obtains 84.0% accuracy and 83.6% macro-F1, compared with 80.0% and 79.2% for reliability-weighted text anchors. At a threshold fixed on validation data, SACR retains 80.8% of test pages at 91.1% selective accuracy. Twenty-run synthetic hOCR perturbation tests show little sensitivity to coordinate jitter and gradual, rather than catastrophic, degradation under token deletion. The method has a median scoring latency of 1.14 ms per page on a CPU and offers an interpretable routing baseline for resource-constrained intelligent document systems.