A Tensor Modeling for Video Rain Streaks Removal Approach Based on the Main Direction of Rain Streaks

Abstract

The algorithms for video rain streaks removal do not properly consider the influence of wind on the main direction of rain streaks. They do not rotate or only perform a rough rotation when rain streaks deviate from the vertical direction, resulting in residual rain patterns or blurred background. Therefore, a sparse tensor model based on the main direction of rain streaks is suggested for video rain streaks removal in this paper. First, the first-order directional derivative (FODD) filter is used to obtain the rain image with the best background suppression effect. Second, we calculate its histogram of oriented gradient (HOG) feature to match the rain streaks image library. The main direction of rain streaks and the rotation angle of the global model are determined by the matching result. Finally, a sparse tensor is constructed with a rotation-angle based regularization term for rain streaks removal. In addition, the tensor nuclear norm (TNN) is replaced with the tensor truncated nuclear norm (T-TNN) to ensure the global low rank of the rain free video. The alternating direction method of multipliers (ADMM) is used to work out the model. The experimental results represent the excellence of the proposed method compared with the baseline methods in terms of the value of peak signal-to-noise ratio (PSNR) and the value of structural similarity (SSIM).