ocr_temp_extractor

Extract temperature readings from a fixed-lab video by selecting a polygon ROI over the on-screen thermometer and running OCR at a fixed time interval. Outputs a CSV with exactly three columns: Medicion,Tiempo,temperatura.

Demo

Features

• Point-based semi-freehand polygon ROI selection on the first frame.
• Time-based sampling using precise timestamp seeks (no frame-by-frame loop).
• Fast preprocessing pipeline (grayscale → resize → denoise → Otsu/adaptive → morphology → deskew with optional sweep).
• OCR via PaddleOCR or Tesseract (select with --model); digits-only post-filtering and parsing.
• Optional preview and interactive assist mode to confirm/correct each value.
• CSV export with header: Medicion,Tiempo,temperatura.

Requirements

• Python 3.10+
• Python packages:
  • Core: opencv-python, numpy, pillow
  • OCR engines (choose one or both):
    • Paddle: paddleocr, paddlepaddle
    • Tesseract: pytesseract (requires system Tesseract binary)
  • CSV uses Python stdlib (csv), no pandas required.

Install PaddleOCR

• CPU (Linux/macOS/Windows):

  pip install --upgrade pip
  pip install "paddlepaddle>=2.5" paddleocr

  If paddlepaddle fails, consult: https://www.paddlepaddle.org.cn/en/install/quick

Install Tesseract OCR

• macOS: brew install tesseract
• Ubuntu/Debian: sudo apt-get update && sudo apt-get install -y tesseract-ocr
• Windows: Install from https://github.com/UB-Mannheim/tesseract/wiki and (optionally) set:
  • setx TESSERACT_CMD "C:\\Program Files\\Tesseract-OCR\\tesseract.exe"

Setup

Create a virtual environment and install dependencies:

python3 -m venv .venv
. .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install --upgrade pip
pip install opencv-python numpy pillow
# Choose OCR engine(s)
pip install "paddlepaddle>=2.5" paddleocr    # for --model paddle
pip install pytesseract                      # for --model tesseract (requires system Tesseract)

Usage

Run the CLI with input video, interval in seconds (float allowed), and output CSV path.

# Basic (PaddleOCR)
python ocr_temp_extractor.py -i lab_cooling.mp4 -t 2.0 -o data.csv --model paddle

# Using Tesseract
python ocr_temp_extractor.py -i lab_cooling.mp4 -t 2.0 -o data.csv --model tesseract

# With preview window (non-blocking)
python ocr_temp_extractor.py --input in.avi --interval 0.5 --output /tmp/out.csv --preview

# Interactive assist mode (review/correct each value)
python ocr_temp_extractor.py -i in.mp4 -t 1 -o out.csv --assist

# Handling tilted screens / tricky lighting
# - Rotate ROI, try larger scale, adaptive thresholding and deskew sweep
python ocr_temp_extractor.py -i in.mp4 -t 1 -o out.csv --rotate 5 --scale 2.5 --adaptive --deskew-sweep

# Edge-preserving denoise (slower but can help 7-seg)
python ocr_temp_extractor.py -i in.mp4 -t 1 -o out.csv --bilateral

# Specify OCR language and allow non-digit chars
# - Paddle: lang like 'en'
# - Tesseract: lang like 'eng'
python ocr_temp_extractor.py -i in.mp4 -t 1 -o out.csv --model paddle --lang en --no-digits-only
python ocr_temp_extractor.py -i in.mp4 -t 1 -o out.csv --model tesseract --lang eng --no-digits-only

ROI Selection Controls

On launching, a window shows the first frame for selecting the temperature display area:

• Modes:
  • Polygon: left-click to add points; p to switch to polygon
  • Rectangle: press b to switch, then left-click once to set the first corner, move the mouse, and left-click again to set the opposite corner (no dragging, avoids Qt panning).
• Common keys: Backspace/u undo, r reset, Enter confirm (auto-closes polygon)
• The polygon/rectangle is rendered with anti-aliased lines and a semi-transparent fill.
• When --rotate is provided, the first frame is rotated for selection so ROI aligns with processed frames.
• Points are stored in normalized coordinates so they scale to any frame size.

Output CSV

• Header (exact): Medicion,Tiempo,temperatura
• Medicion: 1-based index of the measurement
• Tiempo: timestamp in HH:MM:SS from the start (zero‑padded)
• temperatura: parsed float (e.g., 23.500) or empty if OCR fails

Example:

Medicion,Tiempo,temperatura
1,00:00:00,23.500
2,00:00:02,23.375
3,00:00:04,

Preview Window

• Shows the thresholded ROI and overlays the raw OCR text and parsed value.
• Press q to close the preview; processing continues in the background.

Assist Mode

Use --assist to review each sampled timestamp interactively:

• A small window shows the processed ROI and the OCR result.
• Keys:
  • Enter: accept the current value
  • e: edit value (typed in terminal), e.g., 23.45
  • s: set empty (missing/failed OCR)
  • q: quit assist for the remaining samples (continues automatically)
• Note: Assist pauses at each sample; --preview is disabled automatically while assisting.

Advanced Preprocessing Flags

• --scale <float>: Resize before OCR (default 2.0). Larger values can improve digit clarity.
• --adaptive: Use adaptive thresholding (better for uneven illumination) instead of Otsu.
• --bilateral: Use bilateral filter (slower) instead of Gaussian blur; preserves edges.
• --deskew-sweep: Try multiple small angles to refine deskew when the screen is slightly tilted.
• --rotate <deg>: Rotate the ROI by degrees (CCW) before OCR to compensate video tilt.

Notes & Tips

• The tool seeks by time using CAP_PROP_POS_MSEC to avoid drift.
• Minor tilt is corrected via min-area-rect deskew; --deskew-sweep further refines.
• If OCR returns extra characters, the parser extracts only the numeric value.
• If a frame cannot be read near the end of the video, processing stops gracefully and writes what was collected.

Troubleshooting

• PaddleOCR/PaddlePaddle import error: ensure paddlepaddle and paddleocr are installed and match your Python/OS. On some systems you may need a specific wheel from the PaddlePaddle site.
• Tesseract not found: install it and/or set TESSERACT_CMD (Windows path example above).
• Qt plugin error on Wayland: the script sets QT_QPA_PLATFORM=xcb automatically on Linux/Wayland; you can also export QT_QPA_PLATFORM=xcb.
• Video won’t open: check the file path and that the codec is supported by OpenCV/FFmpeg.
• Poor OCR accuracy:
  • Ensure the ROI tightly bounds the digits.
  • Try --scale, --adaptive, --deskew-sweep, or --bilateral.
  • Reduce glare/contrast; the pipeline uses Otsu/adaptive + morphology to clean 7‑seg digits.

Development

Run from source:

. .venv/bin/activate
python ocr_temp_extractor.py -i input.mp4 -t 1.0 -o out.csv --preview

Anomaly Detection

Use the companion script to flag abnormal temperatures and get suggested replacements (linear interpolation or robust local estimate):

python detect_anomalies.py -i out.csv -r anomalies.csv -c corrected.csv

• -r/--report: Writes a CSV listing each anomalous or missing point with the reason and suggested value.
• -c/--corrected: Writes a corrected copy of the input CSV with suggested values filled in for anomalies/missing entries.
• Tuning: --window (default 5), --k-mad (default 3.5), --k-rate (default 5.0).