Interactive defect segmentation in welding radiographic images based on artificial features fusion

Abstract

In recent years, deep learning technology has been used in the defect detection of weld radiographic images with its rapid development. However, there are several questions need to be solved for the wide application of deep learning technology in engineering. First, the lack of prior information due to the lack of large number of training data limits the performance of the model; Secondly, it takes too long for the labeling work of manual discrimination. In addition, when the deep learning prediction is wrong, it is very difficult for human intervention to correct. To solve these problems, a human-computer interaction method for weld defect detection based on HRNet + OCR deep learning model was suggested in this work. In the data set preparation stage, different from the previous processing methods, this paper eliminates the pure background images that do not contain instances, and then not only segmenting the defects in the weld images, but also making different labeling maps for different types of defects and pseudo-defects respectively, solving the problem that the network pays too much attention to the semantic information of the image while ignoring the user interaction when predicting was solved. In the artificial feature extraction phase, based on human experience, the ray image is processed to enhance the non-equilibrium region in the image, especially the non-equilibrium region with small size and weak intensity. Artificial features were integrated into the network, to obtain a stronger and more robust ability to focus and extract the unbalanced areas in the image, this paper proposes to artificial features. The experimental results showed that the best performance of the network can be achieved when the artificial feature convolution kernel with foreground scale of 3 pixels, background scales of 15 pixels and 31 pixels is used in the test data. Through this method, the model can achieve 2.30 and 3.67 in Noc@75 and Noc@80, compared to the model without fusion of artificial features which improves 68.7 % and 64.3 % in Noc@75 and Noc@80, respectively.