| Popis: |
Deep learning-based fire detection models are usually trained offline on static datasets. For continuously increasing heterogeneous sensor data, incremental learning is a resolution to enable incremental updates of models. However, it still encounters the challenge of the stability-plasticity dilemma on cross-domain data. In this paper, we propose a Dynamic Equilibrium Network (DENet) to achieve cross-domain fire detection on heterogeneous images captured by spaceborne, airborne, and terrestrial sensors. It can learn more flame and smoke features from cross-domain data in a balanced manner by imposing dynamic equilibrium constraints on the self-knowledge distillation model with deeply-supervised loss. It adopts learnable loss weights to adaptively adjust the importance of the old and new datasets and focuses on rebalancing the weight bias based on dynamic equilibrium constraints guided by the number of samples. Model performance is evaluated on the Flame And Smoke Detection Dataset (FASDD) consisting of computer vision (CV) and satellite remote sensing (RS) images, as well as the FLAME dataset consisting of Unmanned Aerial Vehicle (UAV) images. Extensive experiments demonstrated that DENet performs better generalization and equalization on CV, RS, and UAV data than other classical incremental learning methods, and significantly narrows the performance gap toward joint training. It outperforms the well-performing PODNet model by +7.71% mAP@0.5 on the old dataset while maintaining the same good performance on the new dataset. In the ablation experiments, our model outperforms the No Bias model by +27.13% mAP@0.5 on the old dataset, and by +2.77% mAP@0.5 on the new dataset. Given its excellent equalization capability on heterogeneous data, DENet can provide effective support for cross-domain fire detection. This study will provide an important reference for collaborative fire warning, emergency response, and government decision-making in a space-air-ground integrated observation network. |
Full Text from ScienceDirect