基于循环一致性对抗网络的室内火焰图像场景迁移

基于深度学习的视频火灾探测模型的训练依赖于大量的正负样本数据,即火灾视频和带有干扰的场景视频。由于很多室内场合禁止点火,导致该场景下的火灾视频样本不足。本文基于生成对抗网络,将其他相似场景下录制的火焰迁移到指定场景,以此增广限制性场合下的火灾视频数据。文中提出将火焰内核预先植入场景使之具备完整的内容信息,再通过添加烟雾和地面反射等风格信息,完成场景与火焰的融合。该方法克服了现有多模态图像转换方法在图像转换过程中因丢失信息而造成的背景失真问题。同时为减少数据采集工作量,采用循环一致性生成对抗网络以解除训练图像必须严格匹配的限制。实验表明,与现有多模态图像转换相比,本文方法可以保证场景中火焰形态的多样性,迁移后的场景具有较高的视觉真实性,所得结果的FID与LPIPS值最小,分别为119.6和0.134 2。

Scenemigration of indoor flame image based on Cycle-Consistent adversarial networks

The training of video fire detection models based on deep learning relies on a large number of positive and negative samples, namely, fire videos and scenario videos involving other disturbances similar to fires. In some instances, the fire video sampled from a scene is insufficient owing to the prohibition of ignition. In this thesis, it was proposed that the flames recorded in other similar scenarios be migrated into the specified scene to increase the data of the fire video in such restricted situations. To complete the content information, a flame kernel was previously implanted into the specified scene, and then style information, such as smoke and ground reflection, were added to fuse the scene and the flame. Our method eliminated the background distortion caused by the loss of information during image translation via the existing multimodal image translation. In addition, Cycle-Consistent adversarial networks were adopted to decrease the dataset quantity and remove the restriction of strict matching of the training images. Compared with other multimodal image-to-image translations, our method ensured that the fire in a scene was diverse, and the migration scene was more visually realistic. The minimum values of FID and LPIPS are achieved, which are 119.6 and 0.134 2, respectively.