Features of building a neural network based on MobileNetV2 models in unmanned aerial vehicle detection tasks

This article delineates the development and calibration of neural network models based on MobileNetV2-SSD (Single Shot MultiBox Detector) and MobileNetV2-SSD FPN-Lite (Feature Pyramid Network) architectures for addressing the detection of unmanned aerial vehicles (UAV). The author conducts a systematic investigation of the structural characteristics underlying the supplementary SSD- and FPN- layers incorporated within these models, aiming to achieve robust detection performance across a spectrum of drone dimensions, as quantified by accuracy and recall metrics. The research validates the documented limitations of existing architectures in detecting small-scale objects within image datasets. By strategically integrating additional layers that assimilate feature representations from the early strata of the base network and generate composite feature maps for detection purposes, the study demonstrates a substantial elevation in mean Average Precision (mAP) by 20 % and an improvement in F1-score balance by 38 % relative to the foundational model. Additionally, the investigation furnishes a comprehensive set of recommendations for hyperparameter optimization to maximize model performance, accompanied by practical guidance regarding training dataset composition and preparation protocols tailored to the particular characteristics of input imagery. Remarkably, the research also revealed an unexpected enhancement in model performance following the quantization process for deployment on edge computing devices, a phenomenon potentially explained by regularization effects inherent to the compilation and quantization pipeline.