| Résumé | The thermal image is an important source of data in the fire safety research area, as it provides temperature information at pixel-level of a region. The combination of temperature value together with precise location information from thermal image coordinates enables a comprehensive and quantitative analysis of the combustion phenomenon of fire. However, it is not always easy to capture and save suitable thermal images for analysis due to several limitations, such as personnel load, hardware capability, and operating requirements. Therefore, it is necessary to have a substitution solution when thermal images cannot be captured in time. Inspired by the success of previous empirical and theoretical study of deep neural networks from deep learning on image-to-image translation tasks, this paper presents a feasibility study on translating RGB vision images to thermal images by a brand-new model of deep neural network. It is called dual-attention generative adversarial network (DAGAN). DAGAN features attention mechanisms proposed by us, which include both foreground and background attention, to improve the output quality for translation to thermal images. DAGAN was trained and validated by image data from fire tests with a different setup, including room fire tests, single item burning tests and open fire tests. Our investigation is based on qualitative and quantitative results that show that the proposed model is consistently superior to other existing image-to-image translation models on both thermal image patterns quality and pixel-level temperature accuracy, which is close to temperature data extracted from native thermal images. Moreover, the results of the feasibility study also demonstrate that the model could be further developed to assist in the analytics and estimation of more complicated flame and fire scenes based only on RGB vision images. |
|---|
