Multi-Wavelet Residual Dense Convolutional Neural Network for Image Denoising

The neural networks with large receptive field show excellent fitting ability and have been successfully applied in image denoising, but with a difficulty to reduce the computational overhead while acquiring good denoising performance. Here we choose a representative of the above networks named multi-wavelet convolutional neural network (MWCNN) as the backbone. To obtain a better tradeoff between the denoising performance and computation speed, we propose to adopt residual dense blocks (RDBs) in each layer of the MWCNN. We call this scheme multi-wavelet residual dense convolutional neural network (MWRDCNN). Benefitting from the applied short-term residual learning strategy, it can increase the learning efficiency. Besides, since we use a hierarchical structure to build our network, the adopted RDBs in different layers are helpful for extracting more object details in different scales. Both horizontal and vertical comparison experiments have been performed to demonstrate the effectiveness of this network in image denoising. The results also show that our MWRDCNN takes much shorter time than other RDB-based networks to extract more features from adjacent layers and is good at handling the images which are badly corrupted by the noise. Thereby, it is a successful attempt to make full use of the advantages of multiple networks without any conflicts.