A small object detection model in aerial images based on CPDD-YOLOv8

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Aerial images contain numerous small objects, posing challenges for accurate detection due to subtle features and background interference.
• Existing object detection models struggle with the scale variation and low visibility of small objects in high-altitude imagery.

Purpose of the Study:

• To develop an improved object detection model, CPDD-YOLOv8, specifically for enhancing the detection of small objects in aerial imagery.
• To integrate novel architectural components to boost the model's ability to capture semantic information and adapt to diverse visual inputs.

Main Methods:

• Proposed CPDD-YOLOv8 incorporates C2fGAM for enhanced semantic understanding, a P2 detection layer for shallow feature extraction, and DSC2f with Dynamic Snake Convolution for adaptive processing.
• The model utilizes a Dynamic Head (DyHead) with integrated attention mechanisms to optimize feature layer weighting.
• Experiments were conducted on the VisDrone2019 dataset, including ablation studies and comparisons with seven other models.

Main Results:

• Ablation experiments confirmed the effectiveness of all proposed modules within CPDD-YOLOv8.
• CPDD-YOLOv8 achieved a higher mean Average Precision (mAP) than seven other models, with mAP@0.5 reaching 41% (a 6.9% improvement over YOLOv8).
• The small object detection rate saw a 13.1% improvement, and model generalizability was validated on WiderPerson, VOC_MASK, and SHWD datasets.

Conclusions:

• CPDD-YOLOv8 demonstrates superior performance in small object detection from aerial images compared to existing methods.
• The novel architectural enhancements effectively address the challenges of detecting small, feature-poor objects in complex aerial scenes.
• The model's robust performance and generalizability indicate its potential for real-world applications in aerial surveillance and analysis.