Research on Lightweight Deep Learning Real-Time Image Recognition Methods for Drone Aerial Photography

Abstract

A prominent contradiction exists between the real-time requirements of drone aerial image recognition and the resource constraints of onboard platforms. Furthermore, the inherent characteristics of aerial images, such as variations in perspective, drastic changes in scale, and the dense distribution of small targets, further intensify the trade-off between model accuracy and inference speed. To address these issues, this paper systematically investigates lightweight deep learning real-time image recognition methods for drone aerial photography. At the backbone network level, this study enhances feature representation capability while maintaining low computational complexity by optimizing the depthwise separable convolution structure and introducing feature reuse and dimensionality reduction mechanisms. At the detector level, this paper designs a hierarchical feature adaptive fusion strategy, which integrates receptive field enhancement and feature refinement techniques to mitigate the issue of missed small target detections. At the deployment level, this study employs structural search based on channel pruning, fixed-point quantization, and operator kernel tuning to achieve the synergistic optimization of model compression and inference acceleration, thereby providing a lightweight solution that balances accuracy and efficiency for real-time drone aerial recognition systems.