Feature Segmentation-based Recognition Technology for Explosive Devices

A novel feature segmentation-based recognition system for explosive devices is proposed, leveraging advanced deep learning techniques to enhance detection accuracy and robustness in complex environments. Effective management of explosive devices is critical for global security, yet existing methods struggle with accuracy and real-time performance. We hope to apply this system to the detection and recognition of conventional explosives or non-conventional explosives, especially highly camouflaged non-conventional explosives. The proposed system integrates Transformer-based Grounding DINO (Grounding by Distillation with Implicit Neural Optimization) with SAM (Segment Anything Model), combining human inputs (e.g., category names or pointers) with CNN (Convolutional Neural Net) and ViTs (Vision Transformers) to improve accuracy, real-time processing, and adaptability. Experiments using visible light and X-ray image datasets of simulated explosive devices demonstrate superior performance of the deep learning model over traditional algorithms in both image segmentation and object recognition tasks. The proposed explosive device identification method significantly enhances accuracy and real-time performance, with great potential for public safety and emergency response applications.