System and Library developed for the inference of an artificial neural network, especially for FaceNet. Made in C and C++ code, it has been tested on an ESP32 microcontroller and raspberry PI.
opencv4.0 or later
- matrix_NHWC_alloc : Allocate memory for NHWC data type
- cmo_lib_free : Free up memory
- cmo_NHWC_l2_normalize : L2 normalization
- cmo_NHWC_MaxPooling : Maxpooling
- cmo_NHWC_AveragePooling: Average Pooling
- cmo_NHWC_conv : Convolution between two volumes
- cmo_NHWC_dense : fully conected
- cmo_NHWC_batch_normalize: batch normalization
- cmo_NHWC_concat: Concatenation
- cmo_NHWC_padding: Applies zero margins to volume
- cmo_NHWC_ActivationRelu: Relu activation function
- Clone the repository.
- In the src folder, download the weights, the link of the FaceNet weights is in the same folder in the file readme2.md
- Compile all the files in this folder with some compiler, it was tested with the cross gcc compiler.
- The program has two options:
- a. Option == 1: Add a new person, it asks for the name and then takes a photo where you can decide whether or not to save the name for that image, if you choose not, the process will be repeated again.
- b. option == 2: Identify the face of the person by comparing each encode that is stored in the database. git clone https://github.com/Itseez/opencv_contrib.git
- Change the detector path on line 112 on the faceNet.cpp file: if(!detector.load("/opt/opencv/data/haarcascades/haarcascade_frontalface_alt.xml")) to their correspondence opencv path.
- In the python_formateo folder, there is a file called formateo.py, this python script converting the weights .h5 (Obtained from TensorFlow) to a file .h (Header C)