Spatially variant scattering-based single-image dust removal in agriculture

Vision systems must deal with various environmental conditions, such as fog, haze, sun, clouds, and snow. A significant challenge in agriculture is the dust raised during soil cultivation. The dust affects image processing systems and algorithms. Like all particles in the atmosphere, dust particles are exposed to light and, therefore, scatter light. Earlier studies have discussed the optics of the atmosphere, identified the scattering process, and developed scene recovery methods. These approaches concentrate on haze removal. However, dust is non-homogeneous and spatially variant. Moreover, dust is raised at different times during the day; thus, the illumination differs. This work addresses the need for a local adaptive scattering model for scene recovery. First, a maximum filter was used to extract the local non-uniform lighting. The final illumination map was further refined with image-guided filtering, which resulted in a blur effect. The blurry map represents the multiple scattering of light through particles. The scene was then recovered on a per-pixel basis. Finally, the proposed method was validated. The validation requires a dataset with images from the same scene with and without dust. A reference-based dataset of nocturnal, dusty, and dust-free images is presented. The results show that the presented approach removes dust with low density but fails if the dust is too dense. Nevertheless, it is visible that dust is significantly reduced, and thus, more information can be extracted. • Reference-based nighttime dataset for vision-based dust removal in agriculture. • The dust raised during tillage in agriculture is a significant challenge. • Spatial variant local illumination estimation with a maximum filter. • Dust is partially removed from the image, even though noise is added. • The code and dataset are open-source available.