The estimation of image noise level is a critical task for image denoising or superresolution reconstruction. Mathematical methods like patch-based or model-based methods suffer from the sensitivity of the selection of homogeneous regions or the selection of a proper statistic model, leading to inaccurate estimation, especially in signal-dependent noise cases such as Rice noise. Ordinary fullyconnected networks often suffer from the over-fitting problem, restricting their usage for realistic images. This article proposes a deep-learning based algorithm by building a deep neural network and it is trained by using the evolutionary genetic algorithm and extreme learning machine (ELM) algorithm extended into Hinton’s dropout framework. By combining the evolutionary genetic algorithm and the proposed extended ELM algorithm, comparative results are obtained, showing higher accuracy and better stability than several state-of-art algorithms.

Convolution, Convolution Neural Network, Deep Learning, Dropout, Estimation, Extreme Learning Machine, Kernel, Mathematical model, Neural networks, Noise level, Noise Level Estimation, Training
IEEE Access
Department of Systems and Computer Engineering

Yang, X. (Xiaohui), Xu, K. (Kaiwei), Xu, S. (Shaoping), & Liu, P. (2018). Image Noise Level Estimation for Rice Noise Based On Extended ELM Neural Network Training Algorithm. IEEE Access. doi:10.1109/ACCESS.2018.2886294