Enhanced UAV target recognition via YOLOv8-SR with optimized super-resolution and hyperparameters

Automatic Target Recognition (ATR) from Unmanned Aerial Vehicle (UAV) images poses significant challenges due to complex landscapes, adverse weather conditions, and low-resolution imagery. This study proposes an enhanced YOLOv8-SR model that leverages optimized super-resolution techniques and hyperparameters to improve detection accuracy and robustness. The model integrates the Inception-NeXt block into the YOLOv8 architecture, enabling detailed feature extraction from UAV images. Pre-processing steps, including contrast enhancement, noise reduction, data augmentation, and normalization, further improve image quality and model performance. Experiments using the VisDrone dataset reveal that the YOLOv8-SR model, optimized with the SGD optimizer, achieves the highest mean Average Precision (mAP@50) of 49.761%, outperforming both Adam and AdamW optimizers. The proposed method significantly enhances target recognition capabilities in UAV images, demonstrating its potential for applications in surveillance, reconnaissance, and security.