MPE-YOLO: enhanced small target detection in aerial imaging.

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Aerial image target detection is crucial for urban planning, traffic monitoring, and disaster assessment.
• Existing algorithms face challenges in accurately detecting small targets within complex aerial environments.

Purpose of the Study:

• To propose an improved YOLOv8-based model, MPE-YOLO, for enhanced aerial image target detection.
• To address limitations in small target recognition and accuracy in complex aerial imagery.

Main Methods:

• Introduced a multilevel feature integrator (MFI) module to improve small target feature representation and reduce information loss during fusion.
• Integrated a perception enhancement convolution (PEC) module to replace traditional layers, boosting fine-grained feature processing.
• Developed an enhanced scope-C2f (ES-C2f) module for improved capture of small target details through channel expansion and multiscale kernels.

Main Results:

• MPE-YOLO demonstrated superior performance on VisDrone, RSOD, and AI-TOD datasets compared to advanced algorithms.
• The model achieved a lightweight structure, indicating enhanced operational efficiency.
• Experimental results confirm MPE-YOLO's potential in improving aerial target detection accuracy.

Conclusions:

• MPE-YOLO offers a significant advancement in aerial image target detection.
• The model effectively enhances both accuracy and efficiency for aerial surveillance applications.
• The proposed modules contribute to overcoming challenges in small object detection in complex environments.