Circular CNN Based Auto-encoding Deep Network for Segmentation of a Welded Circular Seam of Pipeline from an Image Captured by a Camera

It is an important part of an intelligent pipeline nondestructive testing robot to detect the welded circular seams from a frame image which is captured by the robot’s camera. By use of prior knowledge of the contour shape of a welded circular seam, a circular convolution along a circular ring is deigned and an auto-encoding deep learning network is proposed based the circular convolutional neural network this paper. In the stage of training the auto-encoding network, a training frame image is automatically partitioned into concentricity rings. Each ring is visually and easily notated as a welded seam ring or a background ring and then used as a sample to train the network. In the stage of testing, the testing frame is automatically partitioned into concentricity rings according to the sample frame and then the each ring is auto-encoded and recognized if it is welded seam ring. The welded seam is thus coarsely segmented by the circular convolution based auto-encoding deep network. Then OTSU is used to segment out the welded seam from the welded seam ring finely and effectively. The experimental results show that the proposed circular convolution based auto-encoding deep network and segmentation algorithm are effective and efficient.