Research On Deep Learning Based Infrared Thermal Imaging Defect Detection Method

Infrared thermal imaging defect detection technology,with its advantages of high efficiency and speed,no need for contact and easy detection,has rapidly developed into an important branch of the defect detection field and has been highly valued by scholars,and many image processing methods based on this detection method have been derived.From the initial manual observation to determine the defects to a series of defect recognition algorithms based on the spatial domain and frequency domain,which can complete the identification of defects,but the detection model built by these detection algorithms is often optimized to treat the infrared image as an ordinary optical image,with little connection to the detection principle and imaging mechanism,so that a lot of information is lost in the defect recognition,resulting in defect detection accuracy is difficult to It is difficult to further improve the accuracy of defect detection.Therefore,this paper investigates the temperature field sequences of infrared images based on the principle of infrared thermal imaging detection,and combines generative adversarial neural networks to build a more reasonable infrared thermal image enhancement algorithm model to ensure the improvement of defect detection accuracy.The specific work and main research contents of the paper are as follows:(1)Analysis of the difficulties of infrared thermal imaging defect detection in generating adversarial network training.Using the infrared thermal imaging system built in the laboratory as the hardware basis,the difficulties of infrared thermal images in the training process of generating adversarial networks are proposed by analyzing the data characteristics of infrared thermal images of defective specimens.(2)Research on data pre-processing methods for generative adversarial networks.Preprocessing of generator input data is by using PCA algorithm and uniform sampling method,the purpose of eliminating high frequency noise in the thermal time signal is achieved is by linear interpolation of the neighboring thermal time data,the experimental data is expanded to ensure the training effect.Preprocessing of the discriminator input data;An adaptive fast phase extraction algorithm is proposed to normalize and null domain filter the experimentally acquired thermal image sequences frame by frame,and then a single phase map is obtained by using this algorithm,and the corresponding binary image of the thermal image is obtained by edge filling of the phase map to obtain the label image required by the discriminator.(3)A study of infrared thermal imaging defect detection method based on improved generative adversarial network.By combining the characteristics of nonlinearity of Sigmoid function and the variance characteristics of infrared thermal image sequence signal,an infrared thermal imaging defect detection method based on improved generative adversarial network is proposed,and the defect detection of test specimens is completed.A comparative experiment is designed to compare the test accuracy of GAN model with other models and the improved GAN with the original GAN under two defect scenarios.It is proved that the improved GAN has good generalization performance.And by comparing the pixel width of the test results with the real width of the test specimen,it is verified that the proposed method can achieve more accurate detection of defects.In this paper,we analyze the data characteristics of infrared thermal images,propose the training data preprocessing method of GAN and the loss function improvement method of GAN,strengthen the performance of GAN network and reduce the generalization error,improve the accuracy of defect detection,and lay the foundation for the quantitative detection of defects.