This is written for Xiaobai or just getting started with face recognition.
Recently, I got angry when I was looking for a job, and then caused endocrine disorders. I went to the hospital for 10 days. Liaoyang Central Hospital, you can experience the face recognition system they use when you have time.
background
Hospitalized at the hospital was a flow of people relatively large areas, which live in today, tomorrow may be discharged. General patient + one companion. In an environment where the global epidemic has not improved, face recognition has also been used for the entrance control of the hospital.
Model requirement analysis
Analysis of the needs of face recognition system:
- Need real-time recognition
- Need real-time online update-dynamic update of data
- Model has strong anti-noise ability
- Yes, mask recognition. Almost all places need to wear a mask. This function should be available, but the hospital where I live does not. Finally, analyze the research on it.
Model structure analysis
- Face alignment: It should be a small network written by yourself, personally think that SSD is the first choice for reference of face alignment methods;
- Face image: The face of the hospitalized person has changed a lot, either becoming haggard or looking radiant. When training the network, the data should be noisy and whitened;
- Feature extraction network: fast, should be xception or mobilenet v2 similar network. If the device permits, densenet may be added. He is a very good network for feature extraction;
- Calculate distance from database data: There is nothing to say, and the data is updated dynamically.
deploy
1 Local server, it should be one. Why do you say that? When I went to a ward of another department, I recognized who I was, but said I didn't have permission to enter. He is a design bug here. I have identified all the data in the same database for all access control, and are doing the classification of permissions (patients in the department). In this way, a lot of useless calculations will be performed when calculating the features, and the data will be updated.