Open vocabulary detection for concealed object detection in AMMW image

Area of Science:

• Computer Vision
• Artificial Intelligence
• Security Technology

Background:

• Millimeter-wave (MMW) imaging is crucial for security detection.
• Current MMW detectors struggle with novel, unseen object categories.
• Accurate identification of diverse concealed objects presents a significant challenge.

Purpose of the Study:

• To introduce a novel open vocabulary detection algorithm, Open-MMW, for MMW images.
• To enable the recognition of diverse and untrained objects in MMW security detection.
• To address the limitations of closed-set detection in identifying new object types.

Main Methods:

• Adapted the YOLO-World detector framework for MMW image analysis.
• Designed Multi-Scale Convolution and Task-Integrated Block for optimized feature extraction.
• Implemented a Text-Image Interaction Module with attention mechanisms for feature alignment.

Main Results:

• Open-MMW demonstrated significant improvements over baseline models.
• Recall increased by 13.7%, precision by 13.9%, mAP@0.5 by 14.2%, and mAP@[0.5-0.95] by 10.3%.
• Outperformed state-of-the-art multimodal models, showcasing strong zero-shot detection capabilities.

Conclusions:

• Open-MMW effectively addresses the challenge of detecting diverse, unseen objects in MMW images.
• The algorithm offers powerful zero-shot detection capabilities, advancing security applications.
• This work represents the first application of open vocabulary detection to MMW image analysis.