LI-YOLOv8: Lightweight small target detection algorithm for remote sensing images that combines GSConv and PConv

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Small target detection in remote sensing images faces challenges with feature extraction, complex backgrounds, and high computational demands.
• Existing algorithms often struggle with accuracy and efficiency for small objects in diverse remote sensing scenarios.

Purpose of the Study:

• To develop a lightweight and efficient algorithm for small target detection in remote sensing images.
• To enhance feature extraction and reduce computational complexity for improved detection performance.

Main Methods:

• Proposed LI-YOLOv8 algorithm, a modified YOLOv8n, incorporating GSConv, PConv, RFAConv, and a Multi-Scale Attention (EMA) mechanism.
• Replaced SiLU activation with ReLU and integrated a lightweight GP-Detect head.
• Introduced Inner-Wise IoU loss function combining Inner-IoU and Wise-IoU v3.

Main Results:

• Achieved significant improvements in mAP@0.5 (7.6%) and mAP@0.5:0.95 (2.1%) on RSOD dataset compared to YOLOv8.
• Reduced parameters by 10.0% and GFLOPs by 23.2%, demonstrating enhanced efficiency.
• Showcased robust generalization performance across diverse datasets like TinyPerson, LEVIR-ship, brain-tumor, and smoke_fire_1.

Conclusions:

• LI-YOLOv8 offers a substantial advancement in lightweight small target detection for remote sensing.
• The algorithm effectively balances high detection accuracy with reduced computational resources.
• The proposed methods demonstrate strong generalization capabilities for various real-world applications.