On the problem of object recognition on devices with limited resources

The problem of implementing object recognition on microcontrollers with limited resources, particularly the challenges related to their computational power, memory capacity, and energy consumption is addressesed. The use of the compact MobileNet neural network architecture is proposed, which is based on an algorithm with reduced computational load and ensures sufficient performance in resource-constrained environments. This approach enables image classification tasks to be performed on low-cost microcontrollers. Although the technologies discussed are well-known individually, their comprehensive application for specific classification tasks on microcontrollers, such as object recognition, remains underexplored. The article provides a detailed description of all development stages, including data preparation, model parameter tuning, and the use of specialized image scaling methods to reduce computational load. The practical part of the article is devoted to developing a strawberry recognition system to determine their ripeness level using the ESP32 microcontroller. The research results demonstrate the effectiveness of the approach for different applied tasks and confirm the feasibility of integrating computer vision technologies into resource-limited devices. Overall, the work proves that modern machine learning technologies are becoming accessible even to the least powerful hardware platforms, expanding their capabilities and areas of application.