Научен доклад ID 2343 : 2023/1

The existing face recognition algorithms were studied, and the effectiveness of the best of them was substantiated according to a set of criteria, namely: checking accuracy, speed and reliability when working on devices with limited resources, such as embedded devices. Mathematical modeling of the face recognition algorithm designed to work in systems with limited resources was carried out. A study of facial recognition methods was conducted with the aim of choosing the most effective ones for further optimization. It is substantiated that the most effective method of face recognition is mixed convolutional neural networks, namely the method using the FaceNet neural network. A series of experiments was conducted to determine the difference between optimized versions deployed on an embedded device and non-optimized versions. According to the results of experiments, it has been proven that the static quantization model has the best results. Its advantages are: no need to additionally train the model or train it from the beginning, the possibility of optimizing any model according to this principle. It is justified that static quantization eliminates the need to select the loss function and training parameters, as in the case of knowledge distillation or network pruning. It has been proven that from a practical point of view, quantization is the simplest method of optimization, and in terms of accuracy, the experiment proved only minor losses that do not affect the final result.