A reliable anchor regenerative-based transformer model for x-small and dense objects recognition.

Area of Science:

• Computer Vision
• Deep Learning

Background:

• Deep learning models excel at object detection but struggle with x-small and dense objects.
• Existing models suffer from feature extraction limitations and anchor box misalignments, causing score-position discrepancies.

Purpose of the Study:

• To address the limitations in detecting x-small and dense objects.
• To propose a novel feature refinement network using an anchor regenerative-based transformer module.

Main Methods:

• Introduced an anchor regenerative module to generate adaptive anchor scales based on image semantic statistics.
• Integrated a Multi-Head Self-Attention (MHSA) transformer module for in-depth feature map analysis.
• Experimentally validated the model on VisDrone, VOC, and SKU-110K datasets.

Main Results:

• The proposed model demonstrated higher mean Average Precision (mAP), precision, and recall compared to existing methods.
• Achieved superior performance in detecting x-small and dense objects across multiple datasets.
• Showcased excellent fit for VOC and SKU-110K datasets based on accuracy, kappa coefficient, and ROC metrics.

Conclusions:

• The anchor regenerative-based transformer module effectively overcomes limitations in detecting small and dense objects.
• The model's ability to generate dataset-specific anchor scales significantly improves detection accuracy.
• The proposed approach represents a significant advancement in object detection for challenging scenarios.