AL-YOLOv8: A Small Object Detection Algorithm for Remote Sensing Images Based on an Improved YOLOv8s.

Area of Science:

• Computer Vision
• Remote Sensing Technology
• Artificial Intelligence

Background:

• Complex backgrounds and small target sizes in remote sensing imagery lead to false detections.
• Existing algorithms struggle with fusing shallow detail features and deep semantic features for small objects.

Purpose of the Study:

• To develop an enhanced YOLOv8s algorithm, named AL-YOLOv8, for improved small object detection in remote sensing.
• To address challenges posed by complex backgrounds and minute target sizes.

Main Methods:

• The detection head utilizes Adaptive Spatial Feature Fusion (ASFF) for better feature representation.
• Large-Kernel Separate Attention (LSKA) is embedded to enhance small target feature response.
• A novel IFIoU loss function is introduced to mitigate bounding box regression bias.

Main Results:

• AL-YOLOv8 achieved high precision rates of 91.5% (DIOR), 94.2% (RSOD), and 91.8% (NWPU VHR-10).
• mAP@0.5 scores reached 89.8% (DIOR), 96.9% (RSOD), and 92.2% (NWPU VHR-10).
• Consistent improvements over the standard YOLOv8s model were observed.

Conclusions:

• AL-YOLOv8 effectively reduces false detections in small object detection tasks.
• The proposed model significantly enhances detection accuracy for small objects in remote sensing applications.