Weld defect classification using multi level featured unified deep neural network.

In numerous domains, including weld defect classification, deep neural networks (DNN) achieves state-of-the-art results. For weld defect classification, a unified DNN with multi-level characteristics is proposed in this research. To proceed, describe the weld defect features that are being used as inputs to the proposed classification model. The work suggests characteristics based on the intensity difference among the weld defect as well as its own background, in addition to intensity and geometric features. Second, it creates a novel unified DNN, in which the last hidden layer fuses multi-level characteristics from every hidden layer towards determining the class of weld defects completely. It also looks into pre-training and fine-tuning aspects to improve generalisation effectiveness with a short dataset. When compared to other classification methods such as SVM as well as generic DNN models, the framework tends to take full added benefit of features with multi-level, were extracted from the each hidden layer resulting in an excellent performance, with classification accuracy increasing by 3.18 % as well as 4.33 % in the test dataset respectively to attain 98.5% of training accuracy and 93.6% of testing accuracy. [ABSTRACT FROM AUTHOR]