PPM-YOLOv11: Improved YOLOv11n-Based Algorithm for Small-Object Detection in Aerial Images.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Drone aerial imagery presents unique challenges for object detection, including information loss from subsampling, difficulty detecting minute features, and reduced accuracy due to occlusion.
• Existing algorithms struggle with detecting small objects and targets obscured by dense interference.

Purpose of the Study:

• To propose an improved target detection algorithm, PPM-YOLOv11, specifically designed to overcome the limitations of detecting small and occluded objects in drone imagery.
• To enhance the preservation of critical object information and improve detection accuracy in challenging aerial visual conditions.

Main Methods:

• Introduced the C3K2_PPA module, integrating parallelized patch-aware attention with the C3K2 backbone to preserve small object information.
• Developed a multi-scale detection head (P2) for ultra-small objects (4x4 to 8x8 pixels) and added a high-resolution feature layer to the neck network.
• Incorporated the MultiSEAM module to enhance the detection of occluded objects by amplifying features in unobstructed regions and compensating for lost information.

Main Results:

• Achieved 40.9% mAP50 on the VisDrone2019 dataset, a 9.3 percentage point improvement over the baseline YOLOv11n.
• Reached 82.0% mAP50 on the SIMD dataset, surpassing the baseline network by 3.9 percentage points.
• Demonstrated superior performance in detecting small and occluded targets compared to existing methods.

Conclusions:

• PPM-YOLOv11 effectively addresses the challenges of small and occluded object detection in drone aerial imagery.
• The proposed modules and architectural enhancements significantly boost detection accuracy and robustness.
• The algorithm shows strong potential for real-world applications requiring precise aerial surveillance and target identification.