Abstract: Computer recognition of defects in weld testing negative films of chemical equipment and pipelines is a difficult and hot topic in the testing industry in recent years. Due to the limitations of the professional fields, weld testing professionals and image recognition software engineers should learn from each other if they want to do a good job in automatic identification of negative film defects. Testing technicians should understand and master the preliminary theory of convolutional neural network, understand the complete process of convolutional neural network simulating human brain nerves, and conduct in-depth learning to finally identify the defects of weld testing negative films. Similarly, software engineers specializing in image recognition also need to understand and master the expertise of radiographic testing. This paper elaborates on the preliminary theory of convolutional neural network and the process of defect recognition using convolutional neural network. Meanwhile, aiming at the lack of data sources of typical defective images, the data source library is expanded from the perspective of actual radiographic testing based on the on-site practice and many years of experience in film evaluation. Practical verification proves that the above improvements have achieved better results.

Key words: weld testing, deep learning, convolution neural network, data source library, expansion