Deep neural networks (DNNs) have driven the recent advances in fire detection. However, existing methods require large-scale labeled samples to train data-hungry networks, which are difficult to collect and even more laborious to label. This paper applies unsupervised domain adaptation (UDA) to transfer knowledge from a labeled public fire dataset to another unlabeled one in practical application scenarios for the first time. Then, a transfer learning benchmark dataset called Fire-DA is built from public datasets for fire recognition. Next, the Deep Subdomain Adaptation Network (DSAN) and the Dynamic Adversarial Adaptation Network (DAAN) are experimented on Fire-DA to provide a benchmark result for future transfer learning research in fire recognition. Finally, two transfer tasks are built from Fire-DA to two public forest fire datasets, the aerial forest fire dataset FLAME and the large-scale fire dataset FD-dataset containing forest fire scenarios. Compared with traditional handcrafted feature-based methods and supervised CNNs, DSAN reaches 82.5% performance of the optimal supervised CNN on the testing set of FLAME. In addition, DSAN achieves 95.8% and 83.5% recognition accuracy on the testing set and challenging testing set of FD-dataset, which outperform the optimal supervised CNN by 0.5% and 2.6%, respectively. The experimental results demonstrate that DSAN achieves an impressive performance on FLAME and a new state of the art on FD-dataset without accessing their labels during training, a fundamental step toward unsupervised forest fire recognition for industrial applications.
It is important to reduce the danger of collecting flame image data sets by compositing flame images by computer. In this paper, a Global-Local mask Generative Adversarial Network (FGL-GAN) is proposed to address the current status of low quality composite flame images. First, FGL-GAN adopts a hierarchical Global-Local generator structure, to locally render high-quality flame halo and reflection, while also maintaining a consistent global style. Second, FGL-GAN incorporates the fire mask as part of the input of the generation module, which improves the rendering quality of flame halo and reflection. A new data augmentation technique for flame image compositing is used in the network training process to reconstruct the background and reduce the influence of distractors on the network. Finally, FGL-GAN introduces the idea of contrastive learning to speed up network fitting and reduce blurriness in composite images. Comparative experiments show that the images composited by FGL-GAN have achieved better performance in qualitative and quantitative evaluation than mainstream GAN. Ablation study shows the effectiveness of the hierarchical Global-Local generator structure, fire mask, data augmentation, and MONCE loss of FGL-GAN. Therefore, a large number of new flame images can be composited by FGL-GAN, which can provide extensive test data for fire detection equipment, based on deep learning algorithms.
Vision-based forest fire detection systems have significantly advanced through Deep Learning (DL) applications. However, DL-based models typically require large-scale labeled datasets for effective training, where the quality of data annotation is crucial to their performance. To address challenges related to the quality and quantity of labeling, a domain adaptation-based approach called FireDA is proposed for forest fire recognition in scenarios with limited labels. Domain adaptation, a subfield of transfer learning, facilitates the transfer of knowledge from a labeled source domain to an unlabeled target domain. The construction of the source domain FBD is initiated, which includes three common fire scenarios: forest (F), brightness (B), and darkness (D), utilizing publicly available labeled data. Subsequently, a novel algorithm called Neighborhood Aggregation-based 2-Stage Domain Adaptation (NA2SDA) is proposed. This method integrates feature distribution alignment with target domain Proxy Classification Loss (PCL), leveraging a neighborhood aggregation mechanism and a memory bank designed for the unlabeled samples in the target domain. This mechanism calibrates the source classifier and generates more accurate pseudo-labels for the unlabeled sample. Consequently, based on these pseudo-labels, the Local Maximum Mean Discrepancy (LMMD) and the Proxy Classification Loss (PCL) are computed. To validate the efficacy of the proposed method, the publicly available forest fire dataset, FLAME, is employed as the target domain for constructing a transfer learning task. The results demonstrate that our method achieves performance comparable to the supervised Convolutional Neural Network (CNN)-based state-of-the-art (SOTA) method, without requiring access to labels from the FLAME training set. Therefore, our study presents a viable solution for forest fire recognition in scenarios with limited labeling and establishes a high-accuracy benchmark for future research.