ADFireNet: An Anchor-Free Smoke and Fire Detection Network Based on Deformable Convolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Accurate and efficient detection of fire and smoke is crucial for public safety.
• Existing object detection methods often struggle with the complex visual characteristics of fire and smoke.
• Anchor-free methods offer potential advantages but face challenges in label assignment.

Purpose of the Study:

• To propose a novel anchor-free network, ADFireNet, for improved smoke and fire detection.
• To enhance feature extraction capabilities for fire and smoke using deformable convolution.
• To address the label assignment problem in anchor-free networks with pseudo intersection over union.

Main Methods:

• Developed ADFireNet, integrating ResNet with deformable convolution (DCN) in the backbone.
• Employed a feature pyramid network (FPN) in the neck for multi-scale detection.
• Utilized an anchor-free head with pseudo intersection over union (pseudo-IoU) for classification and bounding box regression.

Main Results:

• ADFireNet demonstrated superior accuracy and faster detection speeds on the fire smoke dataset.
• Ablation studies confirmed the significant contributions of DCN and pseudo-IoU to performance.
• The proposed network effectively enhances shape feature extraction for fire and smoke detection.

Conclusions:

• ADFireNet offers a promising solution for real-time and accurate fire and smoke detection.
• The integration of DCN and pseudo-IoU effectively overcomes limitations in current anchor-free detection systems.
• The proposed method shows strong potential for deployment in safety-critical applications.