Neural network architecture for real-time QR code recognition

Abstract

The study investigated modern neural network architectures for efficient real-time recognition of QR codes, which is critical for the development of mobile applications and industrial control systems. The study analysed the features of using light convolutional neural networks optimised for operation on mobile devices with limited computing resources. A modified architecture was proposed that strikes a balance between speed and accuracy when processing a video stream, achieving a recognition rate of 30 frames per second on standard mobile processors. A multi-stage decision-making mechanism based on the Early Stopping Mechanism (ESM) has been developed to optimise image processing. An adaptive filtering method using a median filter and morphological reconstruction was implemented, which substantially improved the quality of input data. The proposed architecture included a specialised preprocessing module and a system of residual-and-excitation blocks to improve recognition efficiency. Experimental studies demonstrated a 12-15% increase in the system`s real-time performance compared to the baseline models when processing a video stream. The system successfully recognised QR codes in poor lighting conditions and non-standard tilt angles with an accuracy of over 92%. A 27% reduction in computational complexity was achieved while maintaining high recognition accuracy. The developed method efficiently processes images with geometric distortions even in conditions of limited resources. The study developed the theoretical foundations of optimising convolutional neural networks for computer vision tasks, offering new approaches to balancing recognition efficiency and accuracy. The practical significance of the study was confirmed by the possibility of direct integration of the developed system into mobile applications and industrial quality control systems, while the proposed optimisation methods can be adapted to a wide range of computer vision tasks on mobile platforms