📝 NEW REPORT OUTLINE: Rakīm AI (رَقِيم)

This outline serves as a blueprint for rewriting the official paper/report for the national competition, aligning it with the verified codebase implementation and removing all outdated technical assumptions.

1. Abstract

Focus: High-level summary of Rakīm AI.
Content: Describe the problem of historical Arabic (Maghrebi) manuscript digitization, the proposed tiered pipeline (segmentation + recognition), dynamic routing based on layout analysis, local offline character-level language models, and semantic post-OCR enrichment. State the final achieved metrics: 7.50% CER on unseen pages and catalog integration of ~50,000 manuscript copies.

2. Problem Statement

Focus: Digital isolation of historical Arabic cultural heritage.
Content:
- Linguistic & Calligraphic Barriers: Standard commercial OCR engines are trained on modern printed fonts and fail completely when confronted with historical handwritten scripts (e.g., Maghrebi calligraphic hands).
- Structural Complexity: Historical manuscripts contain crowded layouts, marginal notes (marginalia / hashiya), page damage, and ink bleed-through.
- Computational Cost: High resource demands of deep learning architectures in academic environments.

3. Objectives

Focus: Engineering a scalable, local-first manuscript processing suite.
Content:
- Build a pipeline that transitions from raw images to editable, search-indexed Arabic texts.
- Develop a dynamic layout router to select optimal model pairs.
- Integrate offline post-OCR language correction to reduce Character Error Rates (CER).
- Deploy semantic features (summarization, indexing, dictionary lookup) and collection matching.

4. System Architecture

Focus: Detailed breakdown of the processing pipeline.
Content:
- Preprocessing & Binarization: Deskewing and Sauvola-based local adaptive thresholding (nlbin).
- Dynamic Layout Routing: Heuristic evaluation of image layout (framed vs. dense vs. open) to select corresponding models.
- Neural Line Segmentation: Extraction of baseline polygons using BLLA (Baseline Layout Analysis) networks.
- Region Classification: Separating main text body (matn) from margins (hashiya) using positional boundaries.
- Neural Text Recognition: Sequence recognition using PyTorch-based convolutional-recurrent networks (CRNN).
- Language Model Shallow Fusion: offline character-level n-gram beam search decoding.
- Performance Layer: SQLite cash-database, FP16 half-precision, and direct multi-worker deployments.

5. Dataset and Model Training

Focus: Fine-tuning methodologies and datasets.
Content:
- Datasets: Use of the RASAM Maghrebi dataset and TariMa historical dataset.
- Training Strategy: Moving away from standard generic printed baselines (reg(96.5)) via low-learning-rate fine-tuning (lrate 0.0001) with learning rate warmup and backbones freezing.
- Data Replay: Balanced line-level (not page-level) replay training to prevent catastrophic forgetting.

6. Model Routing

Focus: Layout-aware model selection.
Content:
- Framed Layout Classifier: Hough-transform-like detection of straight bounding frame borders. Pairs specialized philosophy segmentation (logic_philosophy_v2_seg) with specialized recognition (logic_philosophy_v2_rec).
- Dense Layout Classifier: Color mask ratios (red/black ink) and horizontal ink histograms. Pairs specialized segmentation (logic_philosophy_v2_seg) with general recognition (exp6_best) for handwriting stability.
- General Fallback: Uses seg_best and exp6_best when structural cues are absent.

7. Language Model and Post-processing

Focus: Offline transcription refinement.
Content:
- Shallow Fusion Decoding: Integrating char8_trainval.lm inside CTC beam search decoding to select visually and contextually probable characters.
- Algorithmic Post-Processor: Post-OCR rule engine (PostProcessor.process) to resolve dangling characters, spacing anomalies, and common Arabic letter mergers.

8. AI-Assisted Manuscript Analysis Features

Focus: High-level semantic interpretation.
Content:
- On-Demand LLM Integration: Balancing cloud API latency by calling OpenRouter/Gemini only on user demand.
- Features: Semantic page summary, title suggestions, named entity extraction (scholars, books, dates), contextual definition hovering, and fuzzy Levenshtein search.

9. Experiments and Results

Focus: Factual digital findings (incorporating METRICS_TABLE.md).
Content:
- Compare exp6_best (7.50% unseen CER) with legacy baseline failure.
- Demonstrate language model improvements: 7.48% → 6.79% (RASAM) and 10.27% → 8.81% (TariMa).
- Document layout routing test outcomes (100% correct routing on standard samples).
- Compare CPU vs. GPU latencies: ~14s CPU run vs. ~4s GPU run vs. 0s cash hit.

10. Applications

Focus: Real-world deployments.
Content:
- Academic transcription workspace.
- Local libraries manuscript indexing.
- Duplication matching (cross-collection matching of copies using Fihrist/OpenITI metadata).

11. Limitations

Focus: Objective bottlenecks.
Content:
- Over-Segmentation: Tendency of segmenters to segment margins into small text fragments when borders are faint.
- Catalog Limitation: Local catalog search matches metadata fingerprints, not full-text manuscript content.
- Docker GPU Bounds: Docker container deployment is CPU-based; GPU acceleration requires local PyTorch execution.

12. Future Work

Focus: Roadmap for Rakīm AI.
Content:
- Expanding fine-tuning sets to cover all 8 manuscript families visually identified.
- Training a single robust segmentation model that handles marginalia natively.
- Implementing a local full-text reuse index to search actual transcribed texts against external corpuses.

13. References

Content:
- The RASAM Dataset and TariMa Project.
- OpenITI (Open Islamic Texts Initiative) and FIHRIST (UK Union Catalogue of Islamic Manuscripts).
- Kraken OCR Engine & eScriptorium Framework.
- PyTorch and CUDA developer references.