AMLA-YOLO: Adaptive Multi-dimensional Lightweight Attention for Aerial Small Object Detection

Embedded unmanned aerial vehicles (UAVs) equipped with visual capabilities have become indispensable devices in numerous application domains, including traffic monitoring and urban management. Current methods for UAV-based aerial image detection have partially eased the task of small object detection; however, further advancements are needed to enhance accuracy while concurrently maintaining efficient computational performance. Existing attention mechanisms pre-dominantly employ fixed hyperparameter configurations, making it difficult to adapt to feature maps of varying scales and complexities, which is a key factor constraining real-time aerial image detection performance. This work presents AMLA-YOLO, a real-time detector tailored for small object detection in UAV aerial images, which aims to strike a favorable balance between detection accuracy and computational efficiency. Our method incorporates two lightweight attention modules: the Adaptive SimAM Attention Module (ASim) and the Weighted Lightweight Triplet Attention Module (WLTA), which enhance the feature representation capability of small objects in aerial images. Furthermore, we adopt GhostConv lightweight design and four-scale detection heads to further improve model efficiency. Extensive validation on the VisDrone dataset shows that AMLA-YOLO achieves higher accuracy and faster inference speed compared to various existing real-time detectors.