River floating object detection with transformer model in real time.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Environmental Monitoring

Background:

• Object detection models like DETR and YOLO have advanced significantly.
• RT-DETR improved speed but retained limitations.
• River floating object detection requires efficient and accurate models.

Purpose of the Study:

• To develop a lightweight and efficient object detection model for river floating objects.
• To enhance feature fusion and reduce computational redundancy in object detection.
• To improve the accuracy and real-time performance of river debris detection.

Main Methods:

• Introduced LR-DETR, a lightweight evolution of RT-DETR.
• Incorporated High-level Screening-feature Path Aggregation Network (HS-PAN) for refined feature fusion.
• Utilized Residual Partial Convolutional Network (RPCN) backbone with selective convolutions and residual concepts.
• Integrated RepBlock enhancement with Conv3XCBlock and parameter-free attention mechanism.

Main Results:

• LR-DETR achieved a 5% increase in mean Average Precision (mAP) at IoU 0.5.
• Reduced parameter count by 25.8% and GFLOPs by 22.8% compared to RT-DETR.
• Demonstrated superior performance and adaptability in comparative analyses.
• Showcased significant improvements in real-time river floating object detection.

Conclusions:

• LR-DETR offers a highly efficient and accurate solution for river floating object detection.
• The model's lightweight design and enhanced feature processing are key to its performance.
• LR-DETR shows strong potential for environmental monitoring applications requiring real-time detection.