Rostelecom Speed Recognition

Проект: Автоматическое распознавание скоростей интернета со скриншотов speedtest-сервисов
Цель: Определить, сделан ли скриншот на сервисе Ростелеком (qms.ru) → True, или на другом сервисе → False

YOLOv8n + EasyOCR for extracting download/upload speeds from speedtest screenshots.

---

📁 Structure

speedtest-main/
├── speedtest/
│   ├── main.py                 # Основной скрипт обработки изображений
│   ├── src/
│   │   ├── detector.py         # YOLO детектор областей
│   │   ├── ocr.py              # OCR для распознавания текста
│   │   └── parse.py            # Парсинг числовых значений
│   ├── eval/
│   │   ├── eval.py             # Скрипт оценки качества модели
│   │   └── by_service/         # Тестовые данные по сервисам
│   │       ├── QMS/
│   │       ├── Yandex/
│   │       ├── Speedtest/
│   │       └── Table/
│   ├── data/
│   │   └── processed/yolo_speed/
│   │       └── dataset.yaml    # Конфигурация датасета
│   ├── prepare_dataset.py      # Подготовка датасета (VOC → YOLO)
│   ├── coco_to_voc.py          # Конвертация COCO → VOC
│   └── requirements.txt        # Зависимости
├── weights/                    # Обученные модели (.pt)
└── outputs/                    # Результаты работы скриптов
    ├── results.csv             # CSV с результатами
    └── annotated/              # Изображения с рамками

---

⚙️ Installation

# Перейти в папку проекта
cd speedtest

# Установить зависимости
pip install -r requirements.txt

# Или вручную:
pip install ultralytics easyocr opencv-python pandas openpyxl

---

🚀 Usage

Inference (Creates output folder by default)

# Upload local image
python main.py test.jpg

# Upload through URL
python main.py https://url.com/test.jpg

# Folder (recursive)
python main.py path/images/

# List (one path/URL per line)
python main.py path/images.txt

# ZIP archive
python main.py path/images.zip

# Custom output path + log
python main.py test.jpg --outdir out --log out/log.csv

Output:

• Annotated images: outputs/annotated/
• Terminal: Output dir | Log | Processed: N | Done
• CSV log: outputs/results.csv

CSV columns:

date | filename | download | unit_download | download_unit_source | 
upload | unit_upload | upload_unit_source | det_conf | ocr_conf | 
qms_detected | qms_conf

Field	Description
Speed values	0 is valid, null if not detected
Unit sources	detected (OCR found), copied (from other field), default (fallback to Mbit/s)
qms_detected	True if screenshot is from qms.ru, False otherwise

---

🗂️ Dataset Preparation

Convert COCO → VOC (if needed)

python coco_to_voc.py --input=path/to/coco_annotations --output=path/to/voc_output

Prepare YOLO dataset

python prepare_dataset.py

What it does:

1. Scans eval/AllFotoAndAnnotations for images and VOC XML annotations
2. Converts VOC (.xml) → YOLO format (.txt)
3. Splits data: 70% train / 20% val / 10% test
4. Creates dataset.yaml configuration

Example dataset.yaml:

path: F:/Downloads/speedtest-main/speedtest/data/processed/yolo_speed
train: images/train
val: images/val
test: images/test
nc: 3
names: ['Download', 'Upload', 'qms']

⚠️ Important: Ensure nc (number of classes) and names match your annotations!

---

🎯 Training

yolo detect train \
    data=data/processed/yolo_speed/dataset.yaml \
    model=yolov8n.pt \
    epochs=100 \
    imgsz=640 \
    batch=8 \
    name=qms_final \
    project=runs/detect \
    device=cpu

Parameter	Value	Description
data	dataset.yaml	Dataset configuration path
model	yolov8n.pt	Pretrained model (n/s/m/l/x)
epochs	100	Training epochs
imgsz	640	Image size for training
batch	8	Batch size (reduce if OOM)
device	cpu	cpu or 0 for GPU

Outputs:

• Logs: runs/detect/qms_final/
• Plots: runs/detect/qms_final/results.png
• Best model: runs/detect/qms_final/weights/best.pt

⏱️ Training time: ~3-5 hours on CPU, ~30-60 min on GPU

---

🧪 Testing

# Test on QMS images (should be qms_detected=True)
python main.py eval/by_service/QMS/qms_images --weights="runs/detect/runs/detect/qms_final/weights/best.pt"

# Test on other services (should be qms_detected=False)
python main.py eval/by_service/Yandex/yandex_images --weights="runs/detect/runs/detect/qms_final/weights/best.pt"
python main.py eval/by_service/Speedtest/speedtest_images --weights="runs/detect/runs/detect/qms_final/weights/best.pt"
python main.py eval/by_service/Table/table_images --weights="runs/detect/runs/detect/qms_final/weights/best.pt"

# Test on all images
python main.py eval/AllFotoAndAnnotations --weights="runs/detect/runs/detect/qms_final/weights/best.pt"

Expected results:

Service	qms_detected
QMS (qms.ru)	True (~95%)
Yandex	False
Speedtest.net	False
Table	False

---

📈 Evaluation

cd eval && python eval.py --weights=../runs/detect/runs/detect/qms_final/weights/best.pt

Sample output:

======================================================================
Service               mAP50   mAP50-95  Precision     Recall
======================================================================
QMS                  0.9758     0.7932     0.9446     0.9953
Speedtest            0.8280     0.6383     0.7800     1.0000
Table                0.9950     0.6739     0.9987     1.0000
Yandex               0.9911     0.8111     0.9848     0.9859
======================================================================

Metric	Description	Good value
mAP50	Mean AP at IoU=0.5	> 0.7
mAP50-95	Mean AP at IoU 0.5-0.95	> 0.5
Precision	Fraction of correct detections	> 0.8
Recall	Fraction of objects found	> 0.8

---

🔄 Pipeline

Input → YOLO zones → Crop+Pad → OCR variants → Parse → Annotate+Log

1. Detect Download/Upload boxes (top-1 per class)
2. OCR each crop with preprocessing variants (inverted, threshold)
3. Extract number + unit (Mbit/s, Kbit/s)
4. Fallback: copy unit between fields or default to Mbit/s
5. Save annotated image + append CSV row

Features:

• ✅ Russian+English OCR (EasyOCR)
• ✅ Offline inference
• ✅ URL support
• ✅ Auto-detects CUDA/MPS/CPU
• ✅ QMS service detection (qms_detected flag)

Classes:

• 0 = Download (🔴 red box)
• 1 = Upload (🔵 blue box)
• 2 = qms (🟢 green box, service identifier)

---

❓ Troubleshooting

Problem	Cause	Solution
FileNotFoundError: Weights not found	Wrong model path	Use full path or check runs/detect/runs/detect/
qms_detected=False always	Model not trained on qms class	Check dataset.yaml: nc: 3 and names: [..., 'qms']
Low OCR accuracy	Poor image quality	Use preprocess_variants() in ocr.py
Slow processing	Running on CPU	Use GPU: device=0 or reduce batch
eval.py class mismatch	CLASS_NAMES out of sync	Sync eval.py with dataset.yaml

---

🚀 Quick Start

# 1. Clone repo
git clone <repo-url>
cd speedtest-main/speedtest

# 2. Install dependencies
pip install -r requirements.txt

# 3. Prepare dataset
python prepare_dataset.py

# 4. Train model
yolo detect train data=data/processed/yolo_speed/dataset.yaml model=yolov8n.pt epochs=100 imgsz=640 batch=8 name=qms_final project=runs/detect device=cpu

# 5. Test
python main.py eval/by_service/QMS/qms_images --weights="runs/detect/runs/detect/qms_final/weights/best.pt"

# 6. Evaluate
cd eval && python eval.py --weights=../runs/detect/runs/detect/qms_final/weights/best.pt

---

📞 Contacts & Support

Support for my changes

Field	Value
Author	Андрей
Telegram	@mr_mandarin0
Date	Март 2026

💡 Tip: When encountering errors — check file paths and config sync (dataset.yaml, eval.py, detector.py).

---

Documentation valid for project version from 22.03.2026 🎯