A Weld Seam Dataset and Automatic Detection of Welding Defects Using Convolutional Neural Network
In this paper, we propose a dataset which contains 13006 digitalized x-ray images of welds. And do some preparation work on the original images which can be used to put into the convolutional neural network. Firstly, because of the feature of the input is the original image, but defects of welds are very small in the whole diagram, so that cut up the welds area. Secondly, do some picture preprocessing which includes data enhancement, image regularization, mean subtraction, normalization, etc. Then a model which is built to train and then test the weld images cropped from x-ray images is constructed based on convolutional neutral network. Different from the results ever achieved, this model directly using the feature between pixels and pixels of images without extra extraction of the image feature. Finally, tell the procedure when it comes to train dataset and test dataset, compare the different result of the different image preprocessing, we propose several experiments and results. The results demonstrate that what kind of preprocessing method is better and to do the classification of the picture.
KeywordsWeld image Dataset Deep learning Weld seam detection Convolutional neural network
This research was supported by the National key foundation for exploring scientific instrument of China (Project No. 2013YQ240803).
- 1.Gatys, L.A., Ecker, A. S., Bethge, M.: A neural algorithm of artistic style (2015). https://arxiv.org/abs/1508.06576
- 2.Gatys, L., Ecker, A.S., Bethge, M.: Texture synthesis using convolutional neural networks. In: NIPS 2015 Proceedings of the 28th International Conference on Neural Information Processing Systems, vol. 70, no. 1, pp. 262–270 (2015)Google Scholar
- 4.Raschman, E., Záluský, R., Ďuračková, D.: New digital architecture of CNN for pattern recognition. J. Electr. Eng. 61(4), 222–228 (2010)Google Scholar
- 7.Hassan, J., Awan, A.M., Jalil, A.: Welding defect detection and classification using geometric features. In: 10th International Conference on Frontiers of Information Technology, pp. 139–144 (2012)Google Scholar
- 9.Bai, X., Huang, F.Y., Guo, X.W., Yao, C., Shi, B.G.: Training method and apparatus for convolutional neutral network model. Patent: WO2016155564 (2016)Google Scholar