Improved YOLOv4-tiny based on attention mechanism for skin detection

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Automatic bathing robots require precise skin detection for visually-guided tasks.
• Deep convolutional neural networks (CNNs) offer robust skin detection against environmental variations.
• One-stage object detection algorithms provide real-time performance suitable for practical applications.

Purpose of the Study:

• To enhance skin detection accuracy and efficiency for automatic bathing robots.
• To evaluate the performance of lightweight YOLOv4 variants with attention mechanisms.
• To identify a model balancing detection accuracy and computational cost.

Main Methods:

• Utilized YOLOv4-tiny, a lightweight version of YOLOv4, for skin detection.
• Integrated attention mechanisms (SE, ECA, CBAM) into the YOLOv4-tiny architecture.
• Compared performance against other lightweight backbones (MobileNetV1-V3) integrated with YOLOv4.
• Established a comprehensive evaluation index balancing model size and mean Average Precision (mAP).

Main Results:

• YOLOv4-tiny reduced model size significantly (9.2% of YOLOv4) while retaining substantial mAP (67.3%).
• CBAM and ECA attention modules improved YOLOv4-tiny performance; SE attention decreased it.
• MobileNetVX-YOLOv4 models achieved higher mAP than YOLOv4-tiny variants but had larger file sizes.
• The YOLOv4-tiny model with CBAM demonstrated an optimal balance between size and detection accuracy.

Conclusions:

• Lightweight YOLOv4-tiny with attention mechanisms is effective for robotic skin detection.
• The YOLOv4-tiny model incorporating CBAM offers a practical solution for visually-guided bathing tasks.
• Attention mechanisms can be strategically employed to optimize deep learning models for specific robotic applications.