Resumen
Fire detection employing vision sensors has drawn significant attention within the computer vision community, primarily due to its practicality and utility. Previous research predominantly relied on basic color features, a methodology that has since been surpassed by adopting deep learning models for enhanced accuracy. Nevertheless, the persistence of false alarms and increased computational demands remains challenging. Furthermore, contemporary feed-forward neural networks face difficulties stemming from their initialization and weight allocation processes, often resulting in vanishing-gradient issues that hinder convergence. This investigation recognizes the considerable challenges and introduces the cost-effective Encoded EfficientNet (E-EFNet) model. This model demonstrates exceptional proficiency in fire recognition while concurrently mitigating the incidence of false alarms. E-EFNet leverages the lightweight EfficientNetB0 as a foundational feature extractor, augmented by a series of stacked autoencoders for refined feature extraction before the final classification phase. In contrast to conventional linear connections, E-EFNet adopts dense connections, significantly enhancing its effectiveness in identifying fire-related scenes. We employ a randomized weight initialization strategy to mitigate the vexing problem of vanishing gradients and expedite convergence. Comprehensive evaluation against contemporary state-of-the-art benchmarks reaffirms E-EFNet?s superior recognition capabilities. The proposed model outperformed state-of-the-art approaches in accuracy over the Foggia and Yar datasets by achieving a higher accuracy of 0.31 and 0.40, respectively, and its adaptability for efficient inferencing on edge devices. Our study thoroughly assesses various deep models before ultimately selecting E-EFNet as the optimal solution for these pressing challenges in fire detection.