Research On Defect Recognition Of Ultrasonic Phased Array Image Based On Tensorflow

With the rapid development of science,the ultrasonic phased array detection technology has received more and more attention from the industry.The application of ultrasonic phased array detection technology is inseparable from the analysis of the detection image.It is inevitable that there are errors in the evaluation of defects in the detection image by using artificial methods.Therefore,the use of machine vision technology to identify defects in ultrasonic phased array images is important for improving industrial production efficiency Play a very critical role.Furthermore,in recent years,deep learning has made great progress in image recognition research,and its technology has become increasingly mature.Tensor Flow is a framework for deep learning that is widely recognized in practice.Therefore,based on the Tensor Flow deep learning framework,this paper implements ultrasonic phased array image defect recognition,and explores algorithms including ultrasonic phased array image preprocessing,ultrasonic phased array image segmentation,and ultrasonic phased array image final recognition algorithms.Algorithm and verified by comparative experiments.This article does the following:(1)In terms of ultrasonic phased array image preprocessing,the median filter is a nonlinear filter that has a good effect on image scanning noise and the adaptive histogram equalization algorithm can perfectly save the local information of the image and fuse the median filter And adaptive histogram equalization algorithm for image preprocessing.After experimental comparison and analysis of the classic image preprocessing algorithm,according to the characteristics of the ultrasonic phased array image,median filtering was selected to denoise the image,and then adaptive histogram equalization was selected to enhance it.The ultrasonic phased array image maintains a good color and reduces the problem of color distortion.Its PSNR is 26.6431 and information entropy is 7.0011,which are higher than other preprocessing algorithms such as mean filtering.(2)In terms of ultrasonic phased array image segmentation,in view of the problem that the Grab Cut algorithm based on RGB color space is not ideal for segmenting edges and details,an image combining rough segmentation of HSV color space and Grab Cut segmentation algorithm is proposed.Defective region extraction method.The results of RGB,HSI and HSV three color spaces applied to ultrasonic phased array images were compared experimentally,and the experimental results were evaluated from the aspects of real-time and visual perception.Then,by using color space transformation and Grab Cut algorithm based on HSV space,ultrasonic phased array image defect extraction is used to make up for the shortcomings of Grab Cut algorithm and optimize the performance.)Is about 0.929.(3)In terms of ultrasonic phased array image defect recognition,a seven-layer convolutional neural network model was designed to solve the problems of long training time and large parameter calculation of the Le Net-5 network model.First analyzed the classic neural network model Le Net-5,and then improved it: adding a reconstruction layer,graying the image,and converting it to 4 dimensions to speed up convolution.What the network model "sees" is 4-dimensional data,while retaining the image features,reduces the number of convolution kernel slips and reduces network training time;the use of stacked small convolution kernels instead of large-scale convolution kernels reduces the network 7N(N is the channel of input data Number)parameters;Leaky Re Lu replaces the Sigmoid function in the excitation layer of the Le Net-5 model,enhancing the approximation ability of the deep neural network.Finally,the image preprocessing algorithm,image segmentation algorithm and improved network model proposed in this paper are tested on Caltech101 standard data set and self-built ultrasonic phased array image data set.Firstly,the image preprocessing algorithm and image segmentation algorithm proposed in the article are applied to Caltech101 standard data set and self-built ultrasonic phased array image data set.Then,based on Tensor Flow on Caltech101 standard data set,the designed convolutional neural network structure model can effectively improve the recognition accuracy.The average recognition accuracy of the improved network is 12.6% higher than that of Le Net-5 network.Then,based on Tensor Flow,the self-built ultrasonic phased array image dataset was used to verify the application effect of the network structure model designed in this paper.The improved network applied to the self-built dataset finally obtained a 96.26% recognition accuracy rate.Consider its good application value.Finally,comprehensively compare the application effect of the image processing algorithm proposed in this paper on the two data sets.Compared with directly identifying the original images of the two data sets,the average accuracy of the processed images is increased by 7%,which proves The feasibility and robustness of the processing algorithm.