A propelled multiple fusion Deep Belief Network for weld defects detection

Abstract

With the characteristics including fuzzy edges, high image noise, low pixels and contrast, X-ray images of weld defect are difficult to be effectively recognized. Various well-known deep network is used for improving image recognition performance, so that researchers pay more attention on weld defects detection by using deep network with stack structure. However, such stack structure shows some disadvantages, such as inaccuracy recognition on confusion feature, low uncertainty-handling efficiency, time-consuming and complex computation. In this paper, a propelled multiple fusion Deep Belief Network (PMF-DBN) structure with Fuzzy Classifiers (FC) is created for weld defect classification and recognition. The proposed PMF-DBN enjoy both the ability of DBN neural representation and the of capability of fuzzy representation in order to meet the requirements of variant image feature processing. Meanwhile, instead of time-consuming fine-tuning training, the outputs feature data of each layer is fused in a propelled way, by which effective feature extraction can be achieved. Experiments on weld defects multi-classification demonstrate effectiveness of the PMF-DBN. Compared with the DBN, PMF-DBN has higher recognition accuracy and better fitting performance.