Enhanced Feature Fusion and Transfer Learning for Multi-Format Government Document Classification

Government document digitization faces significant challenges due to diverse formats, degraded quality, and limited annotated data. This paper presents an enhanced feature fusion framework combining convolutional neural networks and transformer architectures for multi-format government document classification. The proposed approach integrates hierarchical visual features with contextual text embeddings via a cross-modal attention mechanism, leveraging progressive transfer learning from general document corpora to specialized government domains. Experimental results on real-world administrative datasets demonstrate classification accuracy improvements of 5.6-8.3 percentage points (pp) over baseline methods, with particular robustness on degraded historical documents. The framework achieves 94.7% accuracy across multiple document formats while maintaining computational efficiency suitable for large-scale deployment in federal and state digitization initiatives.