Image Analysis of the Automatic Welding Defects Detection Based on Deep Learning

Abstract

Automatic detection of welding flaws based on deep learning methods has aroused great interest in the non-destructive testing. However, few studies focus on the characteristics of welding flaws in the X-ray image. This study uses four deep learning models to train and test on a dataset containing 15,194 X-ray images. A hybrid prediction based on OR logic is proposed to avoid the miss detection as much as possible and reduce the miss detection rate to 0.61%, which is state of the art. Quantitative analysis of flaws’ characteristics, including the area, aspect ratio, mean, and variance, suggests the aspect ratios of miss detected flaws are smaller than 2, and the coefficient variances of miss detected flaws are smaller than 0.2. Tracking the critical pixels of X-ray images show that salt noises lead to false alarmed predictions. Error analysis indicates that when using the deep learning method for automatic welding flaws detection, the characteristics of flaws and the factors caused by inappropriate X-ray exposure techniques also should be noted.