An Adaptation and Evaluation for Object Detection of Small Objects Using YOLO8

Detecting small objects in aerial imagery, particularly from UAVs, presents unique challenges due to the reduced size of targets, complex backgrounds, and scale variations. Despite advancements in deep learning and improvements in object detection models, capturing small objects in UAV-captured images remains difficult. Conditions such as object occlusion, varying resolutions, and densely populated environments continue to impact the precision and recall of small object detection, requiring further architectural enhancements to improve performance in these scenarios.In this research, we introduce an enhanced object detection model for small-scale objects in UAV-captured imagery, addressing the unique challenges posed by varying object scales and densely packed environments. Our model incorporates several state-of-the-art improvements to the YOLOv8 architecture. First, we introduce the Multi-Scale Dynamic Attention Mechanism (MSDAM) to optimize feature extraction across multiple scales, enhancing the detection of small objects. Next, we employ a dynamic non-monotonic loss function named Weighted-Intersection over Union (WIoU) to refine bounding box regression, improving localization accuracy, particularly for small, hard-to-detect targets. To further boost performance, Squeeze-and-Excitation (SE) Layers are integrated, adaptively recalibrating feature maps to emphasize the most informative channels, thus enhancing feature representation. Additionally, Soft Non-Maximum Suppression (Soft-NMS) is applied in post-processing, reducing missed detections by preserving overlapping objects in densely populated scenes. Our experiments show significant improvements in precision, recall, and mAP, demonstrating the efficiency of the proposed methodology in accomplishing accurate and computationally efficient object detection for UAV applications.