Capsule Embedded ResNet for Arabic Handwriting Character Recognition

This paper presents novel hybrid architectures for Arabic handwritten character recognition, integrating capsule networks with residual neural networks (ResNets) across various embedding strategies. The proposed Custom Caps-ResNet models explore low-, mid-, high-, and multilevel capsule embeddings to synergize hierarchical feature learning with spatial relationship preservation. Evaluated on four benchmark datasets, the models achieve competitive accuracy—99. 64% on OIHACDB-28 and 94. 14% on Dhad—while consistently reducing test loss by up to 80% on Dhad and 66% on HMBDV1 compared to baselines. These reductions in loss indicate enhanced prediction certainty and improved feature representation. Multilevel and mid-level embeddings perform robustly across diverse script complexities, whereas high-level embeddings excel in semantic abstraction. The variation in dataset performance reveals how capsule networks mitigate challenges in cursive connections, overlaps, and positional character forms. Overall, the integration of capsule embeddings into ResNet hierarchies leads to not only strong accuracy but also significantly more confident predictions—advancing Arabic handwriting recognition toward reliable real-world deployment.