<p>Blurred and degraded text images, commonly encountered in mobile-captured documents, scanned materials, archival records, and historical manuscripts, present significant challenges for accurate text recognition, layout preservation, and semantic restoration. While transformer-based models such as Uformer have made strides in visual restoration, their reliance on fixed-window attention leads to uniform processing across spatial regions, limiting their ability to adapt to content-specific degradations. To address this limitation, we introduce ROBDD+, a Robust and Optimized Blurred Text Document Deblurring framework featuring novel Uformer+, which incorporate tradition Uformer with Dynamic Attention Module (DAM). Unlike traditional attention blocks, DAM incorporates spatial, channel, and filter-wise attention sub-blocks that dynamically reweight feature maps based on their content, enabling fine-grained focus on stroke boundaries, glyph contours, and heavily degraded regions. This content-aware adaptability significantly enhances the clarity and readability of restored text. Beyond deblurring, ROBDD+ incorporates Donut, an OCR-free vision-language model for layout-preserving text extraction and further employs MobiLLaMA as a lightweight language refinement module to recover incomplete text with semantic awareness. To enable mobile deployment, we adopt a two-phase knowledge distillation strategy: (1) Feature-Level Distillation for the deblurring stage using DAM-enhanced features, and (2) Sequence-Level Distillation for text prediction, preserving fidelity in a compressed student model. This dual-stage distillation reduces the model size from 60.29 MB to 23.20 MB (~ 61.52% reduction) without degrading performance. Evaluations on SROIE and NoisyOffice datasets show that ROBDD+ achieves + 12.7% PSNR and + 6.3% SSIM improvements over prior state-of-the-art models, establishing it as a highly effective solution for mobile-friendly document enhancement.</p>