Dual-Path CSDETR: Cascade Stochastic Attention with Object-Centric Priors for High-Accuracy Fire Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Detecting dynamic and amorphous objects like fire and smoke presents significant challenges in computer vision.
• Existing object detection models often struggle with the irregular shapes and rapid changes characteristic of fire and smoke.

Purpose of the Study:

• To develop an advanced object detection model capable of accurately identifying fire and smoke.
• To address the limitations of current methods in modeling irregular object morphologies.

Main Methods:

• Proposing Dual-Path Cascade Stochastic DETR (Dual-Path CSDETR), a novel deep learning architecture.
• Introducing cascade stochastic attention (CSA) for modeling irregular shapes via variational inference.
• Implementing a dual-path architecture for bidirectional feature interaction and enhanced learning.
• Integrating object-centric priors from bounding boxes into decoder layers to refine attention.

Main Results:

• Dual-Path CSDETR achieved a 94% AP50 score on fire and smoke detection tasks.
• The proposed model demonstrated superior performance compared to deterministic baseline models.
• The integration of CSA and the dual-path architecture proved effective in handling amorphous object detection.

Conclusions:

• Dual-Path CSDETR offers a significant advancement in detecting challenging amorphous objects like fire and smoke.
• The model's innovative approach to attention and feature interaction enhances detection efficiency and accuracy.
• This work provides a robust solution for critical applications requiring reliable fire and smoke detection.