Defect Detection And Segmentation Of Industrial CT Image Based On Mask R-CNN

The detection(recognition and location)and segmentation of defects in industrial CT(Computed Tomography)images is arguably an important research topic,in the sense that a large number of workpieces have defects and most of them exist in the form of industrial CT images.At present,the main method to understand the detection and segmentation of workpiece defects is based on DL(Deep Learning).However,existing literature seems to have suggested that there are two main issues in relation to this method:(1)Detection and segmentation are regarded as two independent tasks,which are completed by different algorithms.(2)The result of detection and segmentation has low accuracy when it comes to the defects with fuzzy edge,small size or narrow gray change.To solve the above problems,this dissertation attempts to make improvements on Mask R-CNN(Mask Region with Convolutional Neural Networks)and then applies the improved network to the defects(cracks,bubbles and slags)detection and segmentation of industrial CT images.Based on the target detection network,Mask R-CNN introduces a segmentation branch by using the target segmentation network.The target detection and target segmentation are considered as a whole through shared convolution features,which are combined with each other and complement each other.The main contents of the dissertation are as follows:(1)Extract the defect image.To facilitate the subsequent processing,homomorphic filtering and Laplacian are performed on the image preprocessing to enhance the image.CNN(Convolutional Neural Networks)is used to extract industrial CT images that include the defects in the dissertation,as this network can automatically obtain the features of defects,and is of high recognition accuracy.(2)Mask R-CNN is used to detect and segment the defects in industrial CT images.In the first step,the enhanced basic network extracts the features of the input image to form the feature maps;in the second step,RPN(Region Proposal Networks)chooses and roughly classifies the bounding boxes;in the last step,the network completes the classification,border regression and segmentation of the bounding boxes.The experiments show that the Mask R-CNN achieves good detection and segmentation for most of the defects in this dissertation,but as for the defects with fuzzy edge and small size,there are still very serious problems including false detection,missing detection and low segmentation accuracy.Therefore,this dissertation makes attempts to improve the Mask R-CNN:(1)The top-down feature fusion path in the FPN(Feature Pyramid Networks)is changed into the path combining top-down and bottom-up to enhance the role of low-level features in the feature level.(2)The anchor size is modified according to the characteristics of data set.(3)The size of learning rate isadjusted to get the best experimental results.(4)In training stage,the dropout layer is introduced into the feature extraction model to prevent over fitting.The results suggest that the improved Mask R-CNN can significantly improve the detection accuracy and segmentation accuracy in terms of the defects of low gray contrast,fuzzy edge and small size.(3)Defect measurement.On the basis of the experimental results of the improved Mask R-CNN,perimeter,area and coordinates of the defect segmentation results are measured to master the geometric characteristics and location information of the defect.The experimental results illustrate that improved Mask R-CNN not only realizes the integration of workpiece defect detection and segmentation,but also leads to good results in the accuracy,which is of practical value.