Structure Representation Network and Uncertainty Feedback Learning for Dense Non-uniform Fog Removal

Few existing image defogging or dehazing methods consider dense and non-uniform particle distributions, which usually happen in smoke, dust fog. Dealing with these and/or distributions can be intractable, since fog’s attenuation airlight (or veiling effect) significantly weaken the background scene information input image. To address this problem, we introduce a structure-representation network uncertainty feedback learning. Specifically, extract feature representations from pre-trained Vision Transformer (DINO-ViT) module to recover information. guide our focus on fog areas, then remove accordingly, learning, produces maps, that have higher denser regions, regarded as an attention map represents density uneven distribution. Based map, refines defogged output iteratively. Moreover, handle intractability of estimating atmospheric light colors, exploit grayscale version image, it is less affected by varying colors are possibly present The experimental results demonstrate effectiveness method both quantitatively qualitatively compared state-of-the-art handling smoke.