Adaptive Image Denoising Based on Improved Stacked Sparse Denoising Auto-Encoder

Aiming at the problems that the stacked sparse denoising auto-encoder(SSDA) is difficult to train on image denoising, such as slow convergence rate and poor universality, an adaptive image denoising model based on stacked rectified denoising auto-encoder is proposed. The rectified linear units is used as a network activation function to alleviate the phenomenon of gradient dispersion. Joint training with the residual learning and batch normalization to accelerate convergence speed. In order to solve the problem of noise poor universality of the new model, it is necessary to carry out the multi -channel parallel training, and make full use of the potential data feature extracted by the network to find the optimal channel weights, and learn to predict optimal column weights via training weight prediction model for realizing adaptive image denoising. The experimental results show that the proposed algorithm is not only better than the SSDA in the convergence effect, but also adaptively processing the non-participating training noise, and has better universality, compared with the current methods of BM3D and SSDA.