Noise Benefits For Image Recognition In BP Neural Networks

Under certain conditions,the phenomenon of noise-enhanced signal transmission in nonlinear systems is called stochastic resonance,and the transmission of above-threshold signals by noise-boosting systems is called suprathreshold stochastic resonance.Adding noise to neural networks is an application of suprathreshold stochastic resonance theory,and adding the appropriate amount of noise helps neural network learning during training.After the feedforward neural network is trained by the backpropagation algorithm,it can achieve high accuracy in identifying the MNIST dataset.The main research contents of this paper are as follows:After using coarse grid image feature extraction for MNIST handwritten digits,noise with intensity?_ηand?_?are added to the input layer and output layer of BP neural network to compare the recognition accuracy of the network.Under the same conditions,the image recognition accuracy of BP neural network without noise is 84.2%,and the accuracy increases to 95.25%when Gaussian noise with the best noise intensity?_η=0.7736,?_?=0.2463is added.When the uniform noise with the best noise intensity?_η=0.7445,?_?=0.6693is added,the accuracy is improved by 92.3%;When the Laplace noise with the best noise intensity?_η=0.5565,?_?=0.5609is added,the accuracy is improved by 95.1%,which also verifies that there is stochastic resonance in the image recognition of BP neural network,and Gaussian noise has the best effect,followed by Laplace noise and uniform noise.When the cumulative contribution rate of principal components is 85%,90%,95%and 99%(the feature dimensions are 59,87,154,and 331,respectively),it is found that the recognition accuracy of handwritten digital images with different feature dimensions in the BP neural network with Gaussian noise is different,but they all reach 96%.The dimension reduction of the principal component analysis for the handwritten digital images greatly reduces the system running time,improves the recognition speed,and the recognition effect is better than that of BP neural network without noise,which verifies the existence of noise gain.In addition,compared with the image processing method of coarse mesh feature extraction,the feature extraction method of principal component analysis is more effective for image recognition.Considering that the dimensionality reduction method of principal component analysis sacrifices a certain accuracy while improving the recognition speed,the principal component features and pixel features are fused by the feature weighted combination method of MNIST handwritten digits.When the principal component feature dimensions which are 59,87,154 and 331 are fused with 784-dimensional pixel features,and the image recognition accuracy is compared when the feature dimensions are 843,871,938 and 1115,it is found that the image recognition effect after feature fusion is improved compared with the corresponding dimension when only the principal component is reduced,and the stochastic resonance phenomenon still exists.Among them,when the principal component feature dimension is 87,the accuracy of feature fusion reaches the highest 96.8%,and the running time is the shortest.