Image deraining

Abstract

Image deraining seek to remove rain streaks from rain-filled images. There have been various deep neural network-based image deraining models developed but these models are limited to work smoothly only on devices which have substantial computational capability. This paper implements the lightweight model described in Fu et al. [1] which is usable on devices with low computational capability due to its low number of parameters in the model. We investigate the components (pyramid level, recursive blocks, and loss function) of the model to decide what should be modified. We then tested three modifications namely residual blocks [2], squeeze & excitation [3], and direct extraction of rain streaks. Direct extraction of rain streaks results in the most significant increase of performance. Combining all three modifications yield the best model among implemented models thus far. Implemented models were also tested to determine if they can perform image inpainting. However, even with minor modifications, the models were unable to achieve success in image inpainting.