Learning Enriched Features for Video Denoising with Convolutional Neural Network

Video denoising is of great significance in video processing when shooting conditions are complex such as dynamic scenes and low light. Although existing algorithms have already achieved remarkable denoising performance, the inference time of them is usually impractical for real-time applications. In this paper, we propose a convolutional neural network architecture for video denoising. In contrast to other existing CNN-based methods, our approach utilizes different proportion convolutional kernel numbers in a block for extracting enriched features. Channel attention mechanism is integrated in the network to enhance the denoising performance. The network only needs three contiguous frames and noise map as inputs, which leads to a similar excellent running time to the state-of-the-art. We compare our method with different conventional algorithms VBM4D, VNLB and the state-of-the-art CNN-based method FastDVDnet. Experiment results indicate that our method outputs more convincing results in visual and more robustness than others in both peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) indexes.