Improving the performance of neural network models for use on devices with limited computing resources for household item recognition
Abstract
This article describes the results of scientific research aimed at improving the efficiency of neural network models used in the development of a device that helps blind people find things in their homes. In order to improve the efficiency of neural network models used in object recognition, the results obtained in improving their parameters, such as the reliability of object recognition, image processing speed, and the ability to work on devices with limited computing resources, are presented. In particular, a neural network model specializing in recognizing objects used in homes is created. For this purpose, a specialized model for the required task is created as a result of transfer learning of existing pre-trained models. The models obtained as a result of transfer learning are evaluated based on the model evaluation metrics used in object recognition. The models created as a result of transfer learning are compared by their efficiency according to the evaluation metrics, and the necessary conclusions are made.