This repository contains codes for face recognition project (Automatic attendance system)
I implemented two approaches for face recognition in PYTHON
Project1 folder has implementation of approach 1
Project2 folder has implementation of approach 2
I will recommend to use the Project 2
-
In this approach a Resnet_10 (10 conv layers) with SSD (Single Shot Detector) is used to detect the faces from the image. It is very fast at detecting the faces with descent accuracy.
-
The output of this Resnet_10_SSD (Detected face) is then given to the next CNN VGG_Face_net (16 conv layers). This generates a feature embedding of shape (1,2622).
-
Finally these embeddings are used to train any ML classifier (SVM here).
-
Each Identity (person) can have one or more embeddings. For registering the face, the embeddings are extracted for all the identities and then SVM classifier is trained over them. At the test time, the embedding of the subject (person) is calculated by CNN and SVM predicts the class (indentity of person). Confidence thresholds are used to determine unknow identities.
Below is the project structure:
| File | Description |
|---|---|
| New_User_Images | takes images for a new user and stores into dataset folder |
| Face_descriptor | detects the face and generates feature embedding for any given image |
| New_Embeddings_Extractor | it extracts feature embeddings of newly registered user and saves these embeddings in a pickle file |
| Embeddings | it is the pickle file saved by New_Embeddings_Extractor |
| Classification_Model | it is the ML SVM model that will be trained on new embeddings and saved as Model_pickle |
| Model_pickle | trained ML model |
| Label_encoder.pickle | label encoder saved |
| face_recognition | putting everything in one place |
This approach worked well but did not attain a high accuracy. As a result the 2nd approach was implemented.
In this approach I used dlib library of opencv for face recognition. This was very simple as well as gave a high accuracy results. For more informarion on the library visit link. I also implemented the log file for storing in time and out time for attendace system. The basic working and structure is same as the project 1. But the model backbones are optimized, as a result yeilded better results.
| File | Description |
|---|---|
| recognize_faces.py | takes image as input and returns the label |
| encode_faces.py | creates face encodings and pickles it |
| main.py | runs the application using web camera |
| ipmain.py | runs the application using ip camera |
| dataset | directory that stores images |
How to use:
- Run the script (main.py or ipmain.py)
- Press R to Register a new Face, after that type the name of the person to be registered
- Press P to get Prediction of the current face in the frame.