Research On Defect Detection Algorithms For Cold Cathode X-ray Images

Nondestructive testing is usually achieved by using X-ray imaging systems to detect internal defects in workpieces.Mature and complete imaging systems frequently use a hot cathode as a ray source to emit high-intensity X-rays by emitting hot electrons.Hot cathode Xray sources suffer from defects such as high radiation dose,low radiation conversion efficiency,and the inability to modulate the electron beam rapidly,which seriously affect the performance of the imaging system.Cold Cathode X-Ray sources are becoming a new direction in the development of ray sources during to their unique field emission method,which requires no preheating,low radiation dose,and high conversion efficiency.This paper is based on a cold cathode X-ray imaging inspection system for the detection of welding defects in pipeline welds: a deep learning approach is used to automatically detect different types of defects in cold cathode X-ray pipe weld inspection images.The main work in this paper is as follows:(1)For the original detection image,gaussian and median filters are devoted to removing Gaussian noise and "bright spots",and an exponential enhancement algorithm is accustomed to improve image contrast.The maximum inter-class variance method and morphological operations get used to complete the localization and extraction of the weld area,eliminating the influence of distracting information in the image and facilitating subsequent defect segmentation and identification operations.(2)Comparing the defect segmentation results of traditional segmentation algorithms(Otsu method,Canny edge detection-based segmentation,Watershed algorithm)for cold cathode X-ray inspection images and an analysis of their limitations.In turn,this paper proposes a defect segmentation method based on visual saliency and impulse-coupled neural networks.After experimental testing,the proposed algorithm has a good segmentation effect with high clarity of edge and detail information,and an average segmentation accuracy of95.5% for four defects: cracks,porosity,slag,and imperfections.(3)The VGG16 and ResNet50 convolutional neural network models are used to detect weld defects.The validity of the two models is verified experimentally,and the model performance is evaluated by three metrics: accuracy,recall,and F1 score.The results show that the ResNet50 model outperforms the VGG16 model in all aspects for cold cathode X-ray weld inspection images.To further improve the defect recognition accuracy of the model,this paper incorporates the attention mechanism in the ResNet50 model and changes some of the standard convolutional layers to deeply separable convolutional layers.The improved model can achieve accurate recognition of weld defects,with an accuracy of 95.2% after test set testing,eliminating the uncertainty and complexity of manual feature extraction while ensuring recognition accuracy.