A cascaded group attention mechanism-based object detection algorithm for construction and demolition waste

Area of Science:

• Computer Vision
• Artificial Intelligence
• Waste Management

Background:

• Accurate object detection is vital for effective construction and demolition waste (CDW) management.
• Existing deep learning models struggle with detecting small objects in cluttered construction environments.
• There is a need for efficient algorithms that balance detection accuracy with computational cost.

Purpose of the Study:

• To develop an enhanced YOLOv11 object detection algorithm for improved CDW management.
• To introduce a novel Cascaded Group Attention (CGA) mechanism for better feature extraction.
• To reduce computational and memory overhead while improving detection of small CDW objects.

Main Methods:

• Proposed a transformer backbone incorporating CGA for enhanced long-range dependency modeling and reduced computation.
• Implemented a bidirectional multi-scale interaction module in the neck to fuse high-resolution details with low-resolution semantics.
• Evaluated the algorithm on two datasets, comparing its performance against several YOLOv11-based approaches.

Main Results:

• The proposed algorithm achieved superior performance, with mAP scores of 0.938 and 0.962 on the evaluated datasets.
• The method demonstrated a significant advantage over current state-of-the-art YOLOv11-based algorithms.
• Visualizations confirmed the high accuracy of the model in detecting CDW objects across various scales.

Conclusions:

• The YOLOv11-based algorithm with the novel CGA mechanism effectively enhances the detection of small objects in complex CDW environments.
• The integration of CGA and bidirectional multi-scale interaction modules offers a computationally efficient and accurate solution for CDW object detection.
• This research provides a valuable tool for improving the efficiency and accuracy of construction and demolition waste management.