YOLO - Khmer Text & Math Line Detection (v4.0)

State-of-the-art YOLO models specialized in detecting Khmer text lines and complex mathematical & physics formulas (fractions, square roots, integrals, vectors, limits, matrices, systems of equations) across Cambodian national secondary & high school textbooks (Grades 5 to 12, Mathematics & Physics).

Key Capabilities

• Single-Class Unified Line Detection (text_line): Detects entire horizontal reading lines containing mixed Khmer Unicode text and LaTeX mathematical expressions.
• High Resolution Document Understanding: Optimized for standard high-resolution textbook pages (imgsz=1280).
• Cross-Platform Deployments: Available in both native PyTorch (.pt) and optimized ONNX (.onnx) formats for Server, Desktop, Web (WebAssembly/WebGPU), and Mobile applications.
• Interactive Live Web Studio: Test all model checkpoints in real-time with customizable Confidence, Temper (IoU NMS), in-browser ONNX inference, and full-page viewing at khmer-math-detector Space.

Model Versions & Download Links

All weights are available for direct download from this repository:

Benchmark & Performance Comparison

🛠️ Quickstart Usage

Option 1: Using PyTorch & Ultralytics (Recommended for Desktop/Server)

from ultralytics import YOLO
from huggingface_hub import hf_hub_download

# 1. Download the model from Hugging Face Hub
model_path = hf_hub_download(
    repo_id="krotreaksmey/Yolo_Detect_khmer_math",
    filename="best_khmer_Dectecv2_30e.pt"  # Default v4.0 model
)

# 2. Load model
model = YOLO(model_path)

# 3. Run prediction on a textbook page
# Tip: Use imgsz=1280 for full A4 pages to preserve small subscripts and fractions!
results = model.predict(
    source="path/to/textbook_page.png",
    conf=0.25,
    iou=0.45,
    imgsz=1280
)

# 4. Extract bounding boxes
for r in results:
    boxes = r.boxes.xyxy.cpu().numpy()  # [x1, y1, x2, y2]
    confs = r.boxes.conf.cpu().numpy()
    for box, score in zip(boxes, confs):
        x1, y1, x2, y2 = box.astype(int)
        print(f"Line detected at [{x1}, {y1}, {x2}, {y2}] with confidence: {score:.2f}")

Option 2: Using ONNX Runtime (Cross-Platform / C++ / Python)

import onnxruntime as ort
import numpy as np
from PIL import Image

# Load ONNX model
session = ort.InferenceSession("best_khmer_line_detect_43pct.onnx")

# Load and preprocess image with 640x640 letterbox
im = Image.open("textbook_page.png").convert("RGB")
w, h = im.size
scale = min(640 / w, 640 / h)
nw, nh = int(w * scale), int(h * scale)
im_resized = im.resize((nw, nh), Image.Resampling.BILINEAR)

# Create 640x640 letterbox padded canvas
padded = Image.new("RGB", (640, 640), (114, 114, 114))
pad_x = (640 - nw) // 2
pad_y = (640 - nh) // 2
padded.paste(im_resized, (pad_x, pad_y))

# Normalize to float32 CHW [1, 3, 640, 640]
input_arr = np.array(padded, dtype=np.float32) / 255.0
input_arr = np.transpose(input_arr, (2, 0, 1))[np.newaxis, ...]

# Run inference
outputs = session.run(["output0"], {"images": input_arr})
predictions = outputs[0]  # shape: (1, 5, 8400) -> [cx, cy, w, h, conf]

📚 Training Dataset Distribution (54,407 Samples)

The model training corpus covers the complete Ministry of Education, Youth and Sport (MoEYS) STEM curriculum:

• Mathematics Grade 5: 3,305 images
• Mathematics Grade 6: 3,277 images
• Mathematics Grade 7: 5,276 images
• Mathematics Grade 8: 548 images
• Mathematics Grade 9: 6,869 images
• Mathematics Grade 11 (Advance): 6,669 images
• Mathematics Grade 12 (Basic): 3,745 images
• Mathematics Grade 12 (Advance): 4,475 images
• Physics Grade 11: 7,181 images
• Physics Grade 12: 13,062 images
• Total Unified Images: 54,407 images (Train: 46,245 | Val: 8,162)

Training Visualizations

Related Resources

• Interactive Test Studio: krotreaksmey/khmer-math-detector
• PARSeq OCR Recognition Dataset: krotreaksmey/khmer-math-textbook
• Synthetic Math Formula Dataset: krotreaksmey/synthetic_khmer_math
• Google Colab Test Notebook: Colab_Test_Model.ipynb

Citation

If you use this model or dataset in your research or project, please cite:

@misc{yolo_khmer_math_v3,
  title={YOLO Khmer Text & Math Line Detection Model},
  author={krotreaksmey},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/krotreaksmey/Yolo_Detect_khmer_math}
}