A deep learning-based dynamic deformable adaptive framework for locating the root region of the dynamic flames.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Fire Safety Engineering

Background:

• Traditional optical flame detectors (OFDs) suffer from environmental interference, leading to detection errors.
• Existing deep learning models excel at flame recognition but struggle with dynamic flame source localization.
• Complex environments pose challenges for accurate and robust flame detection systems.

Purpose of the Study:

• To propose a novel Dynamic Deformable Adaptive Framework (DDAF) for accurate dynamic flame detection and localization.
• To address the limitations of existing models in capturing dynamic flame characteristics and root regions.
• To improve the precision and robustness of flame detection in complex industrial and safety scenarios.

Main Methods:

• Introduced Deformable Convolution Network v2 (DCNv2) for adaptive feature extraction of dynamic flames.
• Integrated Context Augmentation Module (CAM) and Dynamic Head (DH) for multi-aspect flame feature analysis.
• Employed Layer-Adaptive Magnitude-based Pruning (LAMP) to optimize model detection speed.
• Developed Inductive Modeling (IM) with coarse- and fine-grained localization for precise flame root delineation.
• Utilized Temporal Consistency-based Detection (TCD) to leverage temporal information for enhanced robustness.

Main Results:

• The DDAF method demonstrated improved AP0.5 by 4.4% compared to classical deep learning methods.
• Achieved significant reductions in model parameters (25.3%) and FLOPs (25.9%), enhancing efficiency.
• Successfully located the flame root region dynamically with high accuracy.
• Experimental results on a custom flame dataset validated the framework's effectiveness.

Conclusions:

• The proposed DDAF framework offers a robust and efficient solution for dynamic flame detection.
• The method significantly enhances the accuracy and speed of flame root localization in complex environments.
• This research extends applicability to critical areas such as industrial safety and combustion process control.