关键词: X射线焊缝缺陷识别, 空洞卷积, 胶囊网络, 空洞分层注意力胶囊网络

Abstract: Due to the complexity and diversity of X-ray images,many traditional methods based on X-ray weld defect detection have poor accuracy and generalization.An X-ray weld defect identification method based on dilated hierarchical attention capsule network(DHACNet)is proposed.DHACNet consists of convolution layer,dilated hierarchical convolution,attention mechanism,and capsule network(CapsNet).Convolution layer is used to extract the convolutional features of images,dilated hierarchical convolution and attention layer are used to extract multi-scale significance features.CapsNet uses capsule layer and dynamic routing algorithm to replace the pooling operation and full connection operation in convolutional neural network(CNN).DHACNet has powerful multi-scale feature extraction capability,which can overcome the drawbacks of CNN,such as focusing only on the local features of the image and the loss of partial image information caused by pooling operation.The experiments are carried out on the constructed X-ray weld defect image set,and its recognition accuracy reached over 96%.Compared with the traditional method the results show that the improved method is effective,and can provide technical support for the X-ray weld defect recognition system.

Key words: X-ray weld defect identification, Dilated convolution, Capsule network, Dilated hierarchical attention capsule network(DHACNet)