Comparative Evaluation of Image Enhancement, Dif-fusion-Based Restoration, and Vision Transformers for Low-Resource Cursive Arabic Handwriting Recognition

Transcribing cursive handwritten Arabic text remains a challenge for computer vision and pattern recognition systems. The presence of small, the right-to-left script, and diagnostically critical sub-character components such as diacritical dots create Uncertainty that many general-purpose recognition pipelines do not resolve well under limited training data. This distinguishes Arabic handwriting recognition from recognition of scripts with simpler character topology and motivates evaluation approaches that account explicitly for both image quality and script-specific structure. This paper reports a comparison of three complementary strategies for improving handwritten-word recognition on the KHATT database under a limited-data, non-pretrained training regime: (i) fuzzy-logic-based local contrast enhancement applied prior to recognition, (ii) denoising diffusion probabilistic model (DDPM)-based image restoration applied prior to recognition, and (iii) a Vision Transformer (ViT) trained end-to-end without external pretraining. All three configurations use an EfficientNetB1-based recognition backbone unless otherwise stated, and are benchmarked against two established references: a convolutional-recurrent network (CRNN) and a pretrained, fine-tuned Transformer-based model (TrOCR). On the evaluated KHATT subset, the proposed fuzzy-logic-enhanced configuration achieved a word accuracy (WA) of 73.29% (CER 20.80%, WER 26.71%, NED 0.184) at 112 frames per second, compared with 70.28% for the unenhanced backbone and 69.59% for the DDPM-restored variant (14 FPS). The CRNN baseline reached 75.10% WA, and the pretrained TrOCR model reached 86.40% WA, while the non-pretrained ViT reached 42.30% WA. By utilizing parameterized fuzzy membership functions alongside a rule-based inference engine, this preprocessing module dynamically adapts local pixel contrast to enhance sub-character details and diacritical marks prior to sequence recognition. We report this comparison, together with an error analysis and an explicit account of the study’s methodological boundaries, as evidence that lightweight fuzzy enhancement strategies remain a highly relevant and computationally efficient option for low-resource Arabic handwriting recognition specifically when large-scale pretraining is not available—a claim stated at this bounded, conditional scope rather than as a general claim that compact models outperform attention-based architectures.