- User upload the images with label which needs to be recognized later by the system
- All the images are processed, the face is detectd by the MTCNN Algorithm and and detected face image will be saved in the database
- For each image(face) 128-d vector representation is generated by pre-trained dlib model or facenet model.
- These encodings are then dumped in .pickle file for future inference purposes.
- For inference, again encodings are generated for face present in the input image.
- Using a threshold value as strictness measure, Euclidean distances are compared through the pickle file and the most closest face is recognised to be that of the user.
- Name and ID of recognised user is displayed.
- Accurate results even in boundary cases:
- Different facial orientation
- Obscured face
- Presence of Spectacles
- Modified facial features(beard)
- RHS image is the image which is saved by the user for later recognition and whose embeddings are stored in .pickle file
- Trained Image of Salman was without Beard, image we given to system to be recognized is having beard, still our system recognises person accurately
- We did not trained virat kohli’s image, and similarity score obtained by our model exceeds our threshold value set, therefore it gives result as unknown person
- here the inference image has different facial orientation and Obscured face, even few people will fail to recognize this image of his struggling days but still our model accurately recognized the face
