Limited-angle CT Reconstruction And Its Application In Defect Detection

Computed tomography(CT)imaging technology has a wide range of applications in industrial non-destructive testing due to its good penetrability and non-destructiveness.However,in some application scenarios,due to the influence of the shape and structure of the detected object and the pursuit of scanning efficiency,only projection data of a limited angle can be obtained,and the reconstruction of limited angle data generally produces artifacts,which seriously affects the detection of material defects.Therefore,there is currently a strong demand in the industry for limited angle CT reconstruction to generate high-quality images.Compared with traditional methods,the limited angle CT reconstruction based on deep learning is superior,but it is very difficult to obtain paired sinogram-reconstructed images for supervised neural network training,and the needs of small-angle limited angle reconstruction are not met.In addition,the defect detection method based on CT images is also a hot spot of current research.The existing detection methods can only detect defects of increased foreign matter.In actual industrial inspection tasks,there are many types of defects that are missing.At this time,conventional defect detection methods will fail.At the same time,in some special detection scenarios,a large amount of defect data cannot be obtained to train the network,which seriously affects the performance of defect detection.Based on the above problems,this thesis systematically conducts the following research:First,a hybrid-domain unsupervised limited angle CT reconstruction algorithm based on Generative Adversarial Networks(GAN)is established.The algorithm first performs image completion in the projection domain,and the completed sinogram is subjected to a filtered back-projection algorithm(FBP)to obtain a reconstructed image with artifacts,and the artifact removal neural network is further used to map the artifact CT image to artifact-free CT image.The comparative experimental results show that the method in this thesis can effectively reduce the limit of projection data to 0°-90°,and the reconstructed images have such indicators as Root Mean Square Error(RMSE),Peak Signal to Noise Ratio(PSNR)and Structural Similarity(SSIM)Compared with the existing methods,it is obviously improved.Then,on the basis of obtaining high-quality limited angle CT reconstruction images,a detection algorithm that can be applied to both missing defects and increased defects is further studied.A CT image-oriented defect detection method is constructed with CSPDarknet53backbone network module,neck network module,network output module and fully connected classification network module.The test results show that the method in this thesis can effectively identify pictures with missing defects.From the results of AP,m AP and other indicators,the method in this thesis is also significantly better than other methods in the performance of identifying increasing defects.Then,aiming at the problems of complex types of small defects and small sample size,which are difficult to detect,a generative adversarial network data augmentation method combining WGAN and capsule network is proposed to improve the small defect detection performance of this method.WGAN fundamentally solves the problem of collapse of the original GAN mode and ensures the training stability.As a discriminator,the capsule network considers the interrelationship and mutual positional relationship between features,improves the discriminator’s ability to reason and understand space,and constrains the network.The resulting defect images are more realistic.The experimental results show that after using the data augmentation method in this chapter,the performance of the defect detection algorithm used for testing has been improved to varying degrees.Finally,because the CT reconstruction method and defect detection method of neural network have problems such as cumbersome training,numerous training parameters and difficult management of the parameter results after training,this chapter designs and implements the defect detection software for limited angle CT reconstruction.Firstly,the overall design of the software is carried out,and based on this,three modules including the limited angle CT reconstruction module,the defect data calibration module and the defect detection module included in the software are implemented.Finally,the software is tested,and the test results show that the software is stable and reliable,and all functions are normal.