A Statistically Validated and Decoding-Aware CNN–Transformer–CTC Framework for Multi-Font Printed Arabic Word Recognition

Printed Arabic Optical Character Recognition (OCR) remains challenging due to complex glyph morphology, typographic variability, and sensitivity to Unicode-preserved evaluation protocols. This work introduces a methodology that explicitly treats decoding strategy and orthographic normalization as primary experimental variables in multi-font Arabic OCR evaluation. A CNN–Transformer encoder trained with Connectionist Temporal Classification (CTC) is employed as a controlled backbone to isolate the effects of inference configuration and text normalization. Through systematic analysis on the APTI benchmark, we demonstrate that decoding policy and diacritic handling significantly influence reported recognition performance. In particular, language-model-guided decoding yields substantial improvements over greedy decoding, while Unicode-preserved evaluation introduces systematic orthographic inflation driven by deterministic diacritic mismatch. These effects are further amplified by strong cross-font variability. The proposed normalization-aware evaluation framework disentangles structural recognition errors from protocol-induced artifacts, providing a more controlled and reproducible basis for Arabic OCR benchmarking.