Publication

Abstract : Handwritten digit recognition is a critical component of postal automation, where real-world envelopes often contain distortions such as skew, smudge, and occlusion. This paper presents a lightweight framework that integrates a Spatial Transformer Network (STN) with depthwise-separable CNN blocks to achieve robustness against geometric variability while remaining computationally efficient. A postal-specific augmentation strategy, including slant, blur, ink bleed, and partial occlusion, further improves generalization to authentic mail streams. The proposed model achieves state-of-the-art accuracy across benchmark datasets and a curated Real-Envelope set while maintaining a compact 5 MB footprint and low inference latency (8 ms per digit). The novelty lies in combining postal-aware augmentation with a lightweight STN-CNN pipeline explicitly designed for deployment in real postal automation systems.