Bayesian Generation Model for Snow Removal
Shilin Li, Jukun Liu · 2022
Snowflakes in the image usually reduce the visibility of the background image and affect the image quality. Nowadays, most of the single image snow removal tasks have made some progress through the design of more complex deep learning models, but most of them use specific datasets, and the generalization ability is not enough. We deal with the snow removal problem from the point of view of the datasets and propose a more effective method to synthesize snow images. The full Bayesian generation model of snow image is established, in which the snow layer is parameterized as a generator and input as some latent variables. To resolve this model, the variational reasoning framework is used to close the expected statistical distribution of snow in a data-driven way. Using the model generator, different and non-repetitive training pairs can be generated automatically, thus enriching and expanding the existing benchmark datasets. In addition, we design a more simple and feasible deep learning snow removal network based on this model. The experimental results show that the generated model can extract the complex snowflake distribution and can effectively improve the current deep learning snow removal performance. the snow removal effect of the snow removal network based on extended datasets is better than that of the existing snow removal models, and has better processing efficiency.