Research On Image Registration And Fusion In UV Imager

High-voltage power equipment will experience a certain degree of insulation degradation under the action of long-term electrical,thermal,mechanical,and other factors,which will seriously threaten the safety of power grid operation.In order to find the defect information of electrical equipment timely and accurately and reduce the probability of failure,the UV imaging detection method is applied to the inspection of power equipment.The highly sensitive UV imager captures the image through the UV camera and visible camera and timely registers and fuses the captured image,which can achieve the purpose of locating the discharge target and discharge point position.Image registration is the premise of image fusion.The higher the registration accuracy,the better the fusion effect can be achieved.The key to registration is to find a suitable geometric transformation model and the optimal spatial transformation parameters.However,the registration of ultraviolet and visible images differs from that of infrared and visual images.The ultraviolet image formed by the discharge of high-voltage electrical equipment captured by the ultraviolet camera has irregular characteristics,and the pixel gray level is quite different from that of the visible image.This paper mainly focuses on the task of UV image and visual image registration and fusion.In this paper,the method of combining convolutional neural network(CNN)and wavelet transform(WT)is used to solve the problem that the characteristics of ultraviolet and visible images in ultraviolet imagers are greatly different,the feature points are not obvious,and the image registration slow speed,etc.In this method,the convolution neural network model can automatically extract the multi-layer features of the image.Combined with the rigid body transformation model,the convolution neural network model has trained many times to output the optimal spatial transformation matrix.The registration of the two images can be completed by moving the UV image according to the calculated spatial transformation parameters,and then the fused image is fused by using the wavelet transform fusion algorithm.The experimental results show that this method can realize the registration and fusion of ultraviolet and visible images quickly and accurately.With the idea of transfer learning,the GoogLeNet model,wavelet transform,and Canny operator are used for UV and visible image registration and fusion.In the transfer learning mode,sharing GoogLeNet network parameters can effectively improve the speed and accuracy of image feature detection.Taking the feature parameters extracted by the GoogLeNet network as the extreme learning machine(ELM)input,the feature extraction and registration of UV and visible images can be realized by continuously training the output spatial transformation matrix.It can avoid the instability of manual feature extraction and effectively improve training efficiency.In order to facilitate the later fault diagnosis and location and determine the details such as discharge position and discharge intensity,the UV image information needs to be retained as much as possible during information fusion.Therefore,the multi-resolution analysis ability of wavelet transform and the edge detection ability of the Canny operator are used to realize the decomposition and reconstruction of image multi-dimensional space.Experimental results show that this method can achieve accurate real-time registration and fusion of UV and visible images to retain all UV information.