Enhanced Object Detection Algorithms in Complex Environments via Improved CycleGAN Data Augmentation and AS-YOLO...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Object detection algorithms struggle in complex environments like poor lighting, adverse weather, and occlusions.
• Existing methods require improvements for robust performance in real-world scenarios.

Purpose of the Study:

• To develop an enhanced object detection framework (AS-YOLO) and data augmentation technique to improve performance in complex environments.
• To address limitations of current object detection models in challenging conditions.

Main Methods:

• Utilized an improved CycleGAN with dual self-attention and spectral normalization for data augmentation.
• Developed the AS-YOLO framework incorporating channel-spatial parallel attention, AFPN structure, and Inner_IoU loss function.

Main Results:

• AS-YOLO showed a 1.5% increase in mAP@0.5 and 0.6% in mAP@0.95 compared to YOLOv8n.
• Data augmentation with style transfer further improved mAP@0.5 by 14.6% and mAP@0.95 by 17.8%.

Conclusions:

• The proposed AS-YOLO framework and CycleGAN-based data augmentation effectively enhance object detection in complex environments.
• The integrated approach demonstrates significant performance gains, offering a robust solution for challenging scenarios.