Multi-scale Feature Fusion for Efficient Video Deraining Algorithm

Video deraining is a critical task in computer vision, as rain streaks degrade the image quality and affect the performance of outdoor vision tasks. Most existing methods are trained on synthetic data, making it difficult to simulate the diverse morphological features of rain in real scenes. For this reason, a video rain removal network with multi-scale feature fusion is proposed. Semi-supervised learning is used to combine synthetic and real data training, and the rain removal network is assisted by a rain streak generation network for rain removal. A multi-scale efficient channel attention mechanism is introduced into the rain removal network for feature fusion to capture spatial features and temporal correlations with less computational resources. Tested on the NTURain dataset, PSNR is improved by 1.73 dB and SSIM by 0.0060 compared to the state-of-the-art method. Experiments demonstrate that it shows good performance in both quantitative metrics and qualitative analysis.