Channel Attention for Fire and Smoke Detection: Impact of Augmentation, Color Spaces, and Adversarial Attacks DOI Creative Commons
Usama Ejaz, Muhammad Ali Hamza, Hyun-chul Kim

и другие.

Sensors, Год журнала: 2025, Номер 25(4), С. 1140 - 1140

Опубликована: Фев. 13, 2025

The prevalence of wildfires presents significant challenges for fire detection systems, particularly in differentiating from complex backgrounds and maintaining reliability under diverse environmental conditions. It is crucial to address these developing sustainable effective systems. In this paper: (i) we introduce a channel-wise attention-based architecture, achieving 95% accuracy demonstrating an improved focus on flame-specific features critical distinguishing backgrounds. Through ablation studies, demonstrate that our attention mechanism provides 3-5% improvement over the baseline state-of-the-art models; (ii) evaluate impact augmentation detection, performance across varied conditions; (iii) comprehensive evaluation color spaces including RGB, Grayscale, HSV, YCbCr analyze reliability; (iv) assessment model vulnerabilities where Fast Gradient Sign Method (FGSM) perturbations significantly performance, reducing 41%. Using Local Interpretable Model-Agnostic Explanations (LIME) visualization techniques, provide insights into decision-making processes both standard adversarial conditions, highlighting important considerations applications.

Язык: Английский

Channel Attention for Fire and Smoke Detection: Impact of Augmentation, Color Spaces, and Adversarial Attacks DOI Creative Commons
Usama Ejaz, Muhammad Ali Hamza, Hyun-chul Kim

и другие.

Sensors, Год журнала: 2025, Номер 25(4), С. 1140 - 1140

Опубликована: Фев. 13, 2025

The prevalence of wildfires presents significant challenges for fire detection systems, particularly in differentiating from complex backgrounds and maintaining reliability under diverse environmental conditions. It is crucial to address these developing sustainable effective systems. In this paper: (i) we introduce a channel-wise attention-based architecture, achieving 95% accuracy demonstrating an improved focus on flame-specific features critical distinguishing backgrounds. Through ablation studies, demonstrate that our attention mechanism provides 3-5% improvement over the baseline state-of-the-art models; (ii) evaluate impact augmentation detection, performance across varied conditions; (iii) comprehensive evaluation color spaces including RGB, Grayscale, HSV, YCbCr analyze reliability; (iv) assessment model vulnerabilities where Fast Gradient Sign Method (FGSM) perturbations significantly performance, reducing 41%. Using Local Interpretable Model-Agnostic Explanations (LIME) visualization techniques, provide insights into decision-making processes both standard adversarial conditions, highlighting important considerations applications.

Язык: Английский

Процитировано

0