Image denoising method of HV insulator damage based on U-Net neural network

Abstract

The noise of the image will lead to large errors in the image processing of high-voltage insulator damage. In order to reduce the impact of noise, a denoising method for high-voltage insulator damage image is designed based on U-Net neural network. The specific method is to define the perception features of the real domain of the HV insulator damage image, to obtain the input and output values in the convolution layer, to construct the excitation function of the convolution layer, to calculate the average value in the selected region, and to obtain the network image feature mapping results. Based on this, displacement changes were acquired by calculating the mean and variance of image data, and multi-scale image features were extracted combined with U-Net neural network. Thus, the image denoising model of high voltage insulator damage is established. Through the continuous optimization algorithm, the value of loss function reaches the predetermined standard, and the denoising of high voltage insulator damage image is finally realized. The experimental results show that as the noise level increases, the peak signal-to-noise(PSNR) ratio obtained by this denoising method shows a decreasing trend. When the noise level is 5 dB and 75 dB, the PSNR values of this denoising method on the color dataset are 27.6 dB and 26.8 dB, respectively. In the grayscale dataset, when the noise level is 75 dB, the PSNR value is 26.9 dB. From the experimental data, it can be seen that the proposed U-Net neural network denoising method is better.