Image Segmentation Of Circle Seam In Cylinder Pipe Inner Surface Based On Deep Autoencoding Network

Automatic search and detection of welds from the frame images captured by the robot optical camera is a basic and important technology for the intelligentization of pipeline inspection robots.Although the traditional image segmentation method has low computational complexity,the segmentation performance is poor in some complex situations,and the segmentation effect obtained does not have good recognizability.In recent years,with the rapid development of deep learning,new ideas have been provided for weld segmentation detection.This article is oriented to the needs of pipeline flaw detection intelligent robots to automatically explore and detect circular welds.It studies the method of fusing the prior knowledge of weld contour shape to deep learning to segment the weld image from the frame image of the inner wall of the circular pipe captured by the robot.This article completes the following research work:(1)According to the prior knowledge that the contour of the weld image is an ellipse or circle,the method of dividing the image ring on the inner wall of the pipe and the matrix representation of the ring image are proposed,so that the weld detection of the entire frame image becomes a smaller scale.The ring image is classified by deep learning,and it makes the labeling of the segmentation target of the frame image sample simple and easy to operate.(2)The autoencoding deep learning network framework of circular convolutional image is constructed,and the convolution operation along the circular is given,so that the positioning of the circular weld in the three-dimensional space from the frame image requires the frame image processing or learning the unscalable image It is guaranteed,and the autoencoding deep learning can be carried out on a smaller image size scale to realize the classification of whether the divided ring is a welded ring,so as to realize the initial segmentation of the weld.(3)Based on the concept of vectorization,the matrix product expressions of convolution operation,pooling operation and nonlinear operation are derived,and the matrix product chain of the forward process and back propagation of the circular convolution image autoencoding deep learning network is obtained expression.And according to the object-oriented method,the convolution class,pooling class,nonlinear class and fully connected class are designed.According to the derived convolution operation,pooling operation and matrix product of nonlinear operation,the front of each layer of deep learning is realized.Propagate operation member functions to and back.(4)For the initially segmented weld ring,the fine segmentation of the weld is realized based on the OTSU algorithm.(5)On the composite video frame of the robot detecting the circular weld in the cylindrical pipe and the video frame captured in actual work,the proposed ring convolution image autoencoding deep learning network was trained and welded.Sewing division.Experimental results show that the autoencoding deep learning network based on circular convolutional image proposed in this paper can achieve effective segmentation of welds.