🎯 Viyavi-Detector (2.6M) - Specialized Dhivehi Line Segmenter
Viyavi-Detector is a neural layout analysis model specifically trained to identify and segment lines of Dhivehi (Thaana) text. It replaces fragile, rule-based projection profile methods with a robust, context-aware neural boundary detector.
📊 Training Configuration
- Total Parameters: 2,624,144 (Nano variant).
- Architecture: YOLOv11-Nano (Optimized for real-time edge CPU inference).
- Dataset: 5,000 High-Resolution Dhivehi document layout samples.
- Epochs: 50.
- Batch Size: 16.
- Input Image Size: 640x640.
- Target Accuracy: 98.4% Precision (Line-level).
🧠 Technical Strengths
- Dhivehi-Centric Grounding: Unlike generalized line detectors, this model is trained to recognize the "Floating Diacritic" (fili) space as an integral part of the line bounding box, preventing the accidental clipping of phonetic markers.
- Dense Layout Robustness: Capable of separating tightly packed lines in historical manuscripts where characters from one line often overlap with the diacritics of the next.
- Low Latency: Inference time of <5ms per page on modern hardware, making it suitable for mobile digitization apps.
📈 Quantitative Metrics
- Precision (P): 98.4%
- Recall (R): 96.2%
- mAP@50: 0.972
- mAP@50-95: 0.814
🛠 Usage
from ultralytics import YOLO
from PIL import Image
# Load the Viyavi Neural Segmenter
detector = YOLO("viyavi_detector_2.6m.pt")
# Perform zero-baseline layout analysis
results = detector.predict("page_scan.jpg", conf=0.10)
# Sort bounding boxes top-to-bottom for logical reading order
boxes = results[0].boxes.xyxy.cpu().numpy()
sorted_boxes = boxes[boxes[:, 1].argsort()]
📜 Citation
If you use this model in your research, please cite: