NUMERICAL METHOD FOR INTEGRATING LAYERS INTO THE ARCHITECTURE OF AN ADAPTIVE ENERGY-EFFICIENT MODEL

This article considers the problem of object recognition in road scenes. The study revealed that the increase in the complexity of neural network models leads to a significant increase in computational costs and energy consumption, which limits their use in mobile devices. To solve the problem, a method is proposed for integrating layers into the architecture of an adaptive energy-efficient model, characterized by dynamic control of the model parameters based on the analysis of the road scene, computing resources, which allows to increase the accuracy of object recognition (road signs, road markings, traffic lights) in road scenes taking into account natural noise (rain, sun glare, fog, snow, twilight) and the energy efficiency of models. The developed method has an advantage over standard YOLOv8 models, providing an optimal balance between recognition accuracy and energy consumption, which is especially important for devices with limited computing resources (smartphones, tablets). Taking into account the addition of stability and optimization layers, adaptive energy-efficient models have modifications: YOLOv8n-AE, YOLOv8s-AE, YOLOv8m-AE, YOLOv8l-AE, YOLOv8x-AE. An example of numerical calculation of the parameters of the «Adaptability» and «Optimization» layers of the road scene is given. The implementation of the method of integrating layers into the architecture of the adaptive energy-efficient model is performed in the application program, which reduces the load on computing resources for use in mobile devices, ensuring accurate analysis of the road scene. The developed models are used in the mobile application «Driver Assistant» for object recognition in real time, ensuring reliable operation of computing resources. In the future, it is advisable to improve the application program, expand the database for additional training of the proposed models.