A unified binary classification network for weld image detection

Quality inspection is one of the important measures to ensure the quality of welding products. In the generation process, there is a lack of weld defects, irregular shape of defects, and unbalanced defect categories, which makes the detection of weld images difficult. At present, the quality inspection of welding products is still carried out by manual inspection of their X-ray images. It is urgent to solve the primary problem of detection that determines whether there is a defect in a weld image. In this paper, we propose a weld image detection method which is converted into a binary classification problem. Our method consists of a feature extraction module and a classification module. The feature extraction module uses residual network to extract feature map based on model parameters learned from large-scale ImageNet dataset, which is fine-tuned on weld images. The classification module combines global average pooling and convolution layers instead of fully collected layer to classify weld images. Experiments show that our proposed method achieves 97.65% true positive rate and 99.85% true negative rate on the weld image dataset. and has better generalization performance in comparison with other methods.