A double-convolution-double-attention Transformer network for aircraft cargo hold fire detection

Abstract

Traditional smoke and gas detection systems in aircraft cargo compartments tend to have high false-alarm rates, and deep learning models reliant on video imagery tend to entail substantial computation. This paper introduces a transfer learning approach, FE-DCDA-Transformer-TL. Color features are used to enhance fire images, so as to improve the recognition of fire smoke and flame targets. The Transformer network is simplified and combined with dual convolution and dual attention mechanism modules. Dual convolution reduces the number of structural parameters of the Transformer network, and dual attention enhances the features of fire smoke and flame. FE-DCDA-Transformer-TL is trained and evaluated on a custom aircraft cargo compartment fire dataset, and tested on a similar dataset. In experiments, the proposed model achieves 97.69% accuracy, 98% precision, 96.7% recall, an F1-score of 97.34%, 0.98 AUC, 3.44G FLOPS, 21.54M Params, and 0.61 FPS. Compared with state-of-the-art methods, the proposed model improves accuracy, precision, and recall by at least 32.91%, 28.60%, and 16.94%, respectively. FE-DCDA-Transformer-TL effectively solves the accuracy problem of aircraft cargo hold fire detection, providing strong support for fire detection.