Real-time face image recognition method using convolutional neural network MobileNetV3

DOI: 10.31673/2412-9070.2025.061207

Abstract

This article describes a developed methodology for real-time face detection and recognition. The methodology is based on an algorithmic pipeline for parallel data processing and exchange between three threads: video camera image acquisition, neural network processing, and recognized image visualization.
The study involved the development of a full real-time data processing pipeline that integratesimage acquisition from a webcam, preprocessing of frames, neural network inference, and visual display of recognition results. The MobileNetV3 model was selected due to its balance between recognition accuracy and computational efficiency. . The main improvement in data processing is achieved through the use of two Convolutional Neural Network (CNN) models with the MobileNetV3 architecture. CNN model #1 was trained for face detection, and CNN model #2 was trained for face recognition. Based on the proposed methodology, a software application was developed using the Python language. Key attention was focused on optimizing the methodology to achieve minimal latency during frame processing.
To overcome the latency inherent to CPU-only systems, a multi-threaded software architecture was designed, consisting of three independent threads responsible for image capture, neural network inference, and visualization. The conducted experiments demonstrated the effectiveness of the developed methodology, which ensures stable operation on a personal computer even without a Graphics Processing Unit (GPU). An average performance of 5.5 frames per second (FPS) with a latency of 0.72 seconds was obtained. The proposed methodology is flexible to changes in the operating system load.

Keywords: facial image recognition; convolutional neural networks; image processing; obileNetV3; Python; machine learning; detection; algorithm; computer vision.