Face Detection and Cropping Pipeline using MTCNN

This repository provides a complete pipeline to detect and crop face regions from images using the MTCNN (Multi-task Cascaded Convolutional Networks) face detector from the facenet_pytorch library.

It is designed specifically for biometric and computer vision research applications where precise face localization and dataset organization are critical. Key features include:

Robust detection of faces captured from various devices (mobile phones, DSLR cameras, etc.) and at multiple distances (e.g., 1m, 10m, 50m).

EXIF orientation correction for mobile-captured images to ensure consistent face alignment.

GPU acceleration support, enabling fast processing of large datasets.

Cropped face output saved as high-quality image patches.

Excel summary generation that logs detailed metadata.

Whether you're working with surveillance footage, academic face datasets, or real-time image capture scenarios, this pipeline provides an efficient and scalable solution for preparing face crops with traceable metadata and repeatable logic.

Goal and Output

This project implements an automated face detection and cropping pipeline using MTCNN from the facenet_pytorch library. It works with high-resolution face datasets (from both Mobile and Camera) and outputs:

• Cropped face images
• An Excel file with detailed metadata:
  • Subject number, image name, camera type, distance
  • Detection status, face box coordinates
  • Image sizes, detection time

Installation

Install the required dependencies: bash pip install facenet_pytorch pip install torch torchvision pip install pandas opencv-python pillow

Important: You must install facenet_pytorch, which includes MTCNN.

Check GPU Availability

To confirm that PyTorch is using GPU: ```python import torch

print("CUDA available:", torch.cuda.isavailable()) if torch.cuda.isavailable(): print("Device count:", torch.cuda.devicecount()) print("Current device index:", torch.cuda.currentdevice()) print("Device name:", torch.cuda.getdevicename(0)) ```

EXIF Orientation Correction

Many mobile images store orientation data in EXIF metadata. This function corrects it before processing: python def load_image_with_correct_orientation(path): img = Image.open(path) try: for orientation in ExifTags.TAGS.keys(): if ExifTags.TAGS[orientation] == 'Orientation': break exif = img._getexif() if exif is not None: orientation_value = exif.get(orientation, None) if orientation_value == 3: img = img.rotate(180, expand=True) elif orientation_value == 6: img = img.rotate(270, expand=True) elif orientation_value == 8: img = img.rotate(90, expand=True) except Exception as e: print(f" EXIF orientation not applied: {e}") return img.convert("RGB")

This step is crucial for face detection to succeed on mobile images.

Directory Setup

Update paths in src/face_detection.py before running: python root_folder = r'path\to\your\image_folder' output_folder = r'path\to\your\output_folder'

Then run: bash python src/face_detection.py

Excel Output Explanation

The output Excel file face_detection_MTCNN.xlsx contains:

| Column | Description | |----------------------|--------------------------------------------| | subject number | Extracted from image filename | | image name | Original image filename | | indoor/outdoor | Parsed from _in_ or _out_ in filename | | device | 'Camera' or 'Mobile' based on filename | | distance | e.g., 1m, 50m, parsed from filename | | category | Live, Spoof, or Props | | detection | Detected or No Detection | | face box coordinate | Bounding box (x1, y1, x2, y2) | | detection time (s) | Time in seconds | | input image size | WxH in pixels | | cropped image size | WxH of cropped face image |

Example image filenames: - 101_in_c_1m_live.JPG - 101_in_c_50m_spoof1_61.JPG - 101_in_c_10m_live_propcat2.JPG

Repository Structure

face-detection-mtcnn-cropping-pipeline/ src/ face_detection.py output/ cropped images face_detection_MTCNN.xlsx requirements.txt LICENSE .gitignore CITATION.cff README.md

License

This project is licensed under the MIT License. See the LICENSE file for details.

Citation

If you use this codebase in your research, please cite: yaml cff-version: 1.2.0 authors: - family-names: "Chowdhury" given-names: "Mohammad Zahir Uddin" affiliation: "Clarkson University" orcid: "https://orcid.org/0009-0009-8194-6669" title: "Face Detection and Cropping Pipeline using MTCNN" date-released: 2025-07-27 license: MIT repository-code: https://github.com/mdzahirdu/face-detection-mtcnn-cropping-pipeline