Small-Target Detection Algorithm Based on STDA-YOLOv8.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Small-target detection is challenging due to network limitations and imbalanced training data, leading to false positives and missed detections.
• Existing datasets often lack sufficient annotations for small objects, hindering model performance.

Purpose of the Study:

• To propose a novel algorithm, STDA-YOLOv8, to enhance small-target detection capabilities.
• To address limitations in feature extraction and data imbalance for improved small-object recognition.

Main Methods:

• Designed a new network architecture incorporating a Contextual Augmentation Module (CAM) with multi-scale dilated convolutions and a Feature Refinement Module (FRM) for adaptive feature fusion.
• Introduced a Copy-Reduce-Paste data augmentation technique to mitigate annotation disparities between small and large objects.
• Conducted ablation and comparative experiments on VisDrone and PASCAL VOC datasets.

Main Results:

• STDA-YOLOv8 achieved 93.5% accuracy on the VisDrone dataset, a 5.3% improvement over YOLOv8.
• Achieved 94.2% accuracy on the PASCAL VOC dataset, a 5.7% improvement over YOLOv8.
• Outperformed mainstream target detection models and specialized small-target detection algorithms like QueryDet.

Conclusions:

• The proposed STDA-YOLOv8 effectively enhances small-target detection performance by improving feature extraction and addressing data imbalance.
• The novel CAM and FRM modules significantly boost detection precision for small targets.
• The Copy-Reduce-Paste augmentation method proves effective in handling annotation disparities, contributing to overall model robustness.