Skip to content

Latest commit

 

History

History
95 lines (65 loc) · 5.1 KB

README.md

File metadata and controls

95 lines (65 loc) · 5.1 KB

FRMN

FRMN is a lightweight Human Face Recognition Model, which is built around a new mobile architecture called MobileNetV2 and the ArcFace Algorithm and is specially designed for embedded devices.

Overview

FRMN Models are built on MobileNetV2.

In our training, we use the ArcFace Algorithm, instead of the traditional Softmax Function, and the Cross-Entropy Loss function.

To reduce the computation complexity, we use images of smaller size (56x56) in our training.

Process of Face Recognition

The steps below list the whole process of a Face Recognition:

  1. Obtain the input images, typecally 320x240 resolution.
  2. Start the Face Detection and obtain the landmark coordinates of the face.
  3. Align the face by using the landmark coordinates and obtain a face image of required size. align_face
  4. Input the aligned face image to the face recognition algorithm and generate a Face ID. get_face_id and recognize_face
  5. Compare the newly generated Face ID against the existing Face IDs and obtain the distance between these two Face IDs (normally in Euclidean distance or Cosine distance).
  6. Determine if the two Face IDs are from a same person by comparing the distance between these two Face IDs and the specified threshold.

Please see the flowchart of face recognition below:

face-recognition-system

API Configuration

  • FACE_REC_THRESHOLD: The face recognition threshold, which is configured as 0.7 and measured in Cosine distance by default.

    • Any two Face IDs with a distance greater than the FACE_REC_THRESHOLD is deemed as the same person.
    • Lower this parameter to increase the recognition rate; and increase this parameter to reduce the rate of false recognition.

  • NOSE_EYE_RATIO_THRES_MIN: The threshold for the minimum ratio of nose-to-left-eye-distance and nose-to-right-eye-distance.

    • Any face image with a ratio lower than the threshold will be deemed as unqualified image and filtered out.
    • Lower this parameter to increase the passing rate of face alignment. Note that this also leads to images with poor quality being used in face recognition.

  • NOSE_EYE_RATIO_THRES_MAX: The threshold for the maximum ratio of nose-to-left-eye-distance and nose-to-right-eye-distance.

    • Any face image with a ratio higher than the threshold will be deemed as unqualified image and filtered out.
    • Increase this parameter to increase the passing rate of face alignment. Note that this also leads to images with poor quality being used in face recognition.

  • FLASH_INFO_FLAG: The ID data flag. ID data exists in the flash whenever this flag is on.

  • FLASH_PARTITION_NAME: Stores the name of the flash partition that stores Face IDs, which shares the same names used in the partitions.csv file.

Recognition Model Selection

5 versions of FRMN models are available by now:

  • FRMN
  • MFN56_1X
  • MFN56_2X
  • MFN56_3X
  • MFN56_4X

Accuracy: MFN56_4X > MFN56_3X > MFN56_2X > MFN56_1X > FRMN

Details

Model FRMN MFN56_1X MFN56_2X MFN56_3X MFN56_4X
Function frmn_q mfn56_42m mfn56_72m mfn56_112m mfn56_156m
Size 1.2MB 1.6MB 2.0MB 2.4MB 3.7MB
Time 363ms 455ms 738ms 1150ms 1540ms

Selection

Models can be selected through idf.py menuconfig or make menuconfig. Select Component config >> ESP-FACE Configuration >> Recognition Model sequentially, you'll see options below.

Recognition model selection

Note: The MFN56_4X model can only be run on the development board with 8MB FLASH

Precautions

Please note the followings when using our Face Recognition Lib:

  • Only RGB images of 56x56 can be used as input images. Using images of other sizes might lower the rate of correct recognition.
  • Call the align_face() function to start face alignment, and only after that, use the aligned images for your face recognition.
  • If you want to use the frmn_q() function to generate Face IDs, please firstly make your input image normalized and fix-pointed by using the transform_frmn_input() function.
  • cos_distance() and euclidean_distance() are used to calculate the distance between two Face IDs, in Cosine distance and Euclidean distance, respectively.
    • For Cosine distance:
      • Range: [-1,1]
      • The larger the Cosine distance between two Face IDs is, the more similar these two Face IDs are.
    • For Euclidean distance:
      • Range: [0,4]
      • The larger the Euclidean distance between two Face IDs is, the more similar these two Face IDs are.
    • Note that, the Cosine distance is used in this example.
  • To store your Face ID in the flash, instead of the RAM, please firstly configure your partitions.csv file.
  • The data stored in ram is in face_id_list format, while in flash is in the format:
  1. 0-31B Info, to check the content in the flash
  2. 32-39B Len, to indicate the number of ids in flash
  3. 40-4095B Reserved
  4. Each id needs 2KB, begins at 4096B