دانلود رایگان ترجمه مقاله توسعه شبکه های عصبی برای کاهش نویز (سال 2010)

توسعه شبکه های عصبی برای کاهش نویز

Development of Neural Networks for Noise Reduction

Abstract

This paper describes the development of neural network models for noise reduction. The networks used to enhance the performance of modeling captured signals by reducing the effect of noise. Both recurrent and multi-layer Backpropagation neural networks models are examined and compared with different training algorithms. The paper presented is to illustrate the effect of training algorithms and network architecture on neural network performance for a given application.

1 Introduction

In physical systems, transmitted signals are usually distributed partially, or sometimes almost completely, by an additive noise from the transmitter, channel, and receiver. The approach investigated in this work is to consider noise reduction as an essentially required process to enhance the estimation process of image reconstruction of the captured signal. Noise reduction is considered as a continuous mapping process of the noisy input data to a noise free output data. The resulted enhanced signal can be applied to the holographic imaging process and improves the performance of the estimated model. Artificial Neural Networks (ANNs) are finding increasing use in noise reduction problems [1, 2, 3, 4, 7, 8, 12, 13, 16, 17], and the main design goal of these Neural Networks (NNs) was to obtain a good approximation for some inputoutput mapping. In addition to obtaining a conventional approximation, NNs are expected to generalize from the given training data. The generalization is to use information that NN learned during training phase in order to synthesize, similar but not identical, inputoutput mapping [11].

In this paper, two different NN architectures are employed. These are Recurrent Neural Networks (RNNs) and MultiLayer Neural Networks (MLNNs). Both networks are trained with five training algorithms. The training functions used are: Gradient descent backpropagation (traingd), gradient descent with momentum backpropagation (traingdm), gradient descent with adaptive lr (learning rate) backprobagation (traingda), gradient descent w/momentum and adaptive lr backpropagation (traingdx), and Leverberg Marquardt backpropagation (trainlm).

The designed NNs are trained with input sequences that are assumed to be a composition of the desired signal plus an additive white Gaussian noise. The networks are expected to learn the noisy training data with the corresponding desired output and generalize the model. This research is an attempt to employ ANN for the enhancement of the measured corrupted signal and reduce the noise. The main contribution includes the following:

• The input training sequences to the designed NNs are assumed to be a composition of the desired signal plus an additive white Gaussian noise. This assumption speeds up the learning process and improves the approximation of the desired model [15].

• The development and comparison of NN architectures for use in noise reduction applications.

• A comparison of modeling performance using multi-layer and recurrent NNs.

• An examination of the relationship between training performance and training speed with the training algorithm used for a given NN architecture.