Stacked Sparse Denoising Auto-Encoder (SSDA) has been successfully applied to image denoising, which is superior to the most of the traditional image denoising algorithms. However, the algorithm has low training convergence speed and poor universality. To address these limitations, we present an adaptive image denoising method based on Deep Rectified Denoising Auto-Encoder. To reduce training difficulty and speed up convergence, the rectified linear units (ReLu) is used as activation function of the network, and batch normalization (BN) is used to normalizes the input data of the middle layer in network. Different from the previous image denoising model, the proposed model is trained to learn the mapping relationship between noisy image and noise by residual learning. To overcome the problem that the new model cannot effectively deal with noise not seen in training, we perform adaptive training of multiple channels on the improved model to obtain the optimal channel weights and jointly output the denoised image. The experimental results show that the proposed algorithm can not only outperform SSDA in the convergence speed, but also adaptively remove the noise that is not seen in training.
展开▼
