Design and Implement of Compressed Domain Framework for Handwritten Document Translation and Text Summarization

Abstract

Compressed-domain document understanding is gaining relevance as handwritten archives are typically captured, stored, and transmitted in compressed form. Yet, conventional optical character recognition (OCR) pipelines still resort to full image decompression prior to recognition and downstream language processing. This decode-first approach increases latency, memory footprint, and implementation complexity, especially for large-scale archival systems, edge deployments and privacy-sensitive environments where minimal data bloat is preferred. We propose COMP-TransSummNet, an end-to-end compressed-domain framework for handwritten document translation and summarization, which directly processes JPEG transform coefficients, performs compressed-domain handwritten text recognition, and then feeds confidence-aware recognition outputs into neural machine translation (NMT) and abstractive summarization. The core algorithm, Compressed Coefficient Sequence Encoding with a Dual-Attention Transformer (CCSE-DAT), involves three complementary mechanisms: a learned frequency-gating over discrete cosine transform (DCT) coefficients that downweights quantization-sensitive components, a dual-attention encoder that combines coefficient-space evidence with layout and quantization metadata, and a confidence-propagation channel that carries token-level confidence estimates into translation and summarization. Finally, on compressed handwritten document settings derived from IAM-style and cBAD-style benchmarks, the proposed framework achieves a word accuracy of 92.0% and a character error rate of 9.1% in the compressed domain, as well as a BLEU score of 30.6 and a ROUGE-L score of 44.8 for the downstream language tasks, while decreasing per-image latency to 0.24 s and relative memory consumption to 0.61 compared to a decompress-then-process baseline. These results show that coefficient-space learning preserves discriminative handwritten structure, that confidence-aware propagation reduces cascade errors in the recognition-translation-summarization pipeline, and that compressed-domain processing provides a desirable accuracy efficiency trade-off for multilingual document intelligence.