Abstract: Smoke at the forest fire scene is prone to obscure the information of fire field elements, thus seriously interfering with the implementation of remote sensing technology tools such as fire point localization. In this paper, an improved CycleGAN network algorithm for non-homogeneous smoke filtering in fire scene is proposed. The algorithm combines a wavelet transform branch with a knowledge distillation branch containing Res2Net module and attention module to form a generator, introduces a PatchGAN network as a discriminator, and at the same time adds perceptual loss and mapping loss functions on the basis of CycleGAN. The algorithm realizes the recovery of fire details and can effectively remove forest fire image artifacts. In order to verify the effectiveness of the model, the effect of the algorithm is compared with the existing de-fogging algorithms based on the international typical dataset NH-HAZE and the mesoscale point-burning experimental dataset. The results show that the algorithm's de-fogging effect is significantly improved compared with the existing model (compared with the second place, the PSNR and SSIM values are improved by at least 2.4 dB and 2.07 dB,0.02 and 0.15, respectively), which is able to improve the quality of remote sensing monitoring of forest fires, and it can provide richer and more valuable key information about the fire scene for forest fire fighting decision-making.

Key words: forest fire, non-homogeneous smoke, smoke filtering, improving CycleGAN networks, discrete wavelet transform, knowledge adaptation