YOLO11x-text-detection was fine-tuned for Handwritten Russian Dialectic Text Detection in dialectological cards. The model was trained for 100 epochs with a batch size of 32 using dual NVIDIA T4 GPUs. Fifteen layers were frozen during training as part of the transfer learning process. The entire training took approximately 7 minutes.

What is a dialectological text?

Linguists at NaRFU go on dialectological expeditions to different villages of Arkhangelsk region. The dialogs with locals are transcribed into notebooks and the examples of a dialect words and an example of its usage is written on cards. The dialectological text is a text that conveys linguistic features using special symbols like acutes, apostrophes etc.

Example of a card:

Example Usage

python

# Load libraries
import cv2
from ultralytics import YOLO
from pathlib import Path
import matplotlib.pyplot as plt
from huggingface_hub import hf_hub_download


# Download model
model_path = hf_hub_download(repo_id="Daniil-Domino/yolo11x-dialectic", filename="model.pt")

# Load model
model = YOLO(model_path)

# Inference
image_path = "/path/to/image"
image = cv2.imread(image_path).copy()
output = model.predict(image, conf=0.3)

# Draw bounding boxes
out_image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
for data in output[0].boxes.data.tolist():
    xmin, ymin, xmax, ymax, _, _ = map(int, data)        
    cv2.rectangle(out_image, (xmin, ymin), (xmax, ymax), color=(0, 0, 255), thickness=3)

# Display result
plt.figure(figsize=(15, 10))
plt.imshow(out_image)
plt.axis('off')
plt.show()

Metrics

Below are the key evaluation metrics on the validation set:

Precision: 0.940
Recall: 0.924
mAP50: 0.972
mAP50-95: 0.656