Research On Image Defect Detection Of Industrial CT Based On Faster R-CNN

Industrial CT is an advanced nondestructive testing technology,which can reflect its internal condition without destroying the structure of the workpiece.At present,the defect detection algorithm of industrial CT image(detection of bubble,slag and crack)are mainly based on the traditional methods,these methods basically follow the idea of “manual design features + classifier”.There are two main problems in the traditional defect detection algorithm:(1)the way of region selection is using a sliding window to slide on the image continuously,and this strategy is not targeted,the time complexity is high and the window is redundant;(2)the defect features of manual design are not robust to the variety of changes,and are only suitable for the detection of specific defects.In light of above problems,this thesis combines the latest technology in the field of image recognition,an industrial CT image defect detection method based on deep learning is studied.The convolution network is used to automatically extract the feature of the defect target,then location and identification the defect target according to the feature extracted;in the end,on the basis of location,the defect area is segmented and measured to obtain the spatial location of the defect.The main research work is as follows:(1)Defect detection.In the thesis,a defect detection algorithm based on Faster R-CNN(Regions with Convolutional Neural Networks Features)is studied to solve the problem that defect features needs to be manually designed in traditional defect detection algorithms.Firstly,the defect images are annotated with rectangle boxes(Ground Truth box,GT box);then the feature is extracted using convolution neural network,and Region Proposal Networks(RPN)is applied for regional proposal;finally,judging the category of the proposal region and regress the bounding box of the target.Through the analysis of the results of the experiment,it is found that Faster R-CNN has a poor detection effect on the defect area with small size of GT box,the false detection rate and missing detection rate are relatively high.Therefore,the Faster R-CNN network structure has been improved:(1)The characteristics of different convolution layer fusion as a new feature;(2)The scale and the aspect ratio of anchor are reselected according to the length width ratio of the labeled GT box and the size of the GT area;(3)The introduction of Online Hard Example Mining(OHEM)algorithm after the Ro I(Region of Interest)pooling layer,and it is mainly aim at high loss region in the training process;(4)The sets is enhanced by laplace operator and homomorphic filtering before training.The comparison experiment shows that the improved network detection result is better and the detection precision is higher.(2)Defect segmentation and measurement.In order to make a quantitative analysis for defects detection result,the defect area after Faster R-CNN localization is reconstructed by morphology in the first place;then the Otsu double threshold method is used to transform the reconstructed region,and the location area is segmented step by step which according to whether the identification results of the location area is slag or not(because the gray distribution of cracks and bubbles is opposite to the slag);the parameters such as area,circumference,center of mass and others are measured in final.The location,identification,segmentation and measurement of industrial CT image defects are completed through the above steps.In the defect detection phase,the intelligence level is high,which avoids the problem of poor robustness of manual design defects feature,while maintains high recognition accuracy and reducing time cost.A new method for defect detection of industrial CT images is provided in the thesis.