This project dives into the world of facial recognition technology, leveraging the power of the DeepFace library to perform real-time face verification.
The verify_face function is the heart of the facial recognition logic. Here's what it does:
-
Functionality: It takes a frame from the video feed, a reference image, and a result container as inputs. Using the DeepFace library, it compares the face in the frame with the face in the reference image to verify if they match.
-
Implementation: It's designed to run in a separate thread every 30 frames to ensure smooth real-time operation without slowing down the video feed. If there's no face detected or if the comparison fails, it safely sets the result to False to avoid errors.
The main function orchestrates the entire process:
-
Initialization: It sets up the video capture, loads the reference images, and initializes variables for tracking results.
-
Main Loop: Runs in a continuous loop, capturing frames from the camera. Every 30 frames, it triggers the verify_face function in a separate thread for performance efficiency.
-
Display: Based on the verification result, it overlays "MATCH!" in green or "NO MATCH!" in red on the video feed, providing real-time feedback to the user.
-
Exit: The loop breaks when the user presses 'q', ensuring a graceful shutdown by releasing the camera and closing all OpenCV windows.
-
Import: from deepface import DeepFace brings the powerful DeepFace library into our project.
-
Role: DeepFace is crucial for this project as it provides pre-trained models for facial recognition, specifically the VGG-Face model which I used for our verification process. It simplifies the complex task of facial recognition by abstracting away the deep learning intricacies, allowing us to focus on application logic.
One of the significant challenges I encountered was related to the DeepFace dependency, particularly with downloading the vgg_face_weights.h5 file:
-
Download Issues: Initially, the automatic download process for the model weights failed multiple times, likely due to network issues or permissions. This caused repeated attempts to download the file, which I observed in the terminal output.
-
Solution: I resolved this by manually downloading the weights file and placing it in the project directory, ensuring the script could find and load it properly. This workaround was necessary since DeepFace didn't accept custom paths for the weights directly in the function call at the time.
A major visual enhancement added to the project is the dynamic facial recognition mesh:
-
Implementation: Using Google's MediaPipe Face Mesh, the program now detects and draws approximately 400 facial landmarks on the live webcam feed.
-
Dynamic Color Coding:
- The facial mesh appears green when a match is detected ("Access Granted").
- The mesh appears red when no match is detected ("Access Denied").
-
Purpose: This gives a more futuristic and professional facial recognition effect, while also offering immediate visual feedback to the user.
-
Technical Details: The program processes each frame to detect facial landmarks, and dynamically changes the color of the drawn mesh depending on the real-time verification status.
-
Recognizes Faces: When I looked directly into the camera, the system recognized my face and displayed "MATCH!" in green, confirming a successful verification.
-
Testing Images: In the face_recognition_results directory, testing images from one up until four were when the program did not function correctly. Testing images from five to seven are examples of how the program was able to recognize and unrecognize my face based on the reference_image.jpg.
I learned not just about implementing AI in practical applications but also about managing dependencies, handling real-time data, and dealing with the nuances of model performance. By extending the project with facial landmark visualization and voice feedback, the system now feels much closer to real-world facial recognition solutions. While it works well for direct face recognition, there's still room for further growth — especially in making it even more robust for profile views and different lighting conditions.