Defect detection of vaccine glass tubes based on semantic segmentation

Glass tubes for vaccines have strict production requirements, and large-scale production is difficult to ensure their high quality. Minor damage or defects may seriously affect the quality of vaccines. In the actual task of vaccine glass tubes defect detection, there are many kinds of defects, and the defect forms of the same kind are also very different. Traditional manual detection is time-consuming, laborious and expensive. The traditional digital image processing algorithm is also poor in the field of defect detection. This paper proposes a DD-STDC net (STDC net for defect detection) based on semantic segmentation. Multi-scale input branches are introduced on the basis of the original STDC net, so that the model can learn the regional relationship between different scales through attention. For all kinds of irregular defects, deformable convolution is used in the backbone to capture the edge details of irregular objects more easily. The attention mechanism is embedded in the backbone network to strengthen the attention to the effective spatial location and the effective semantic information channel. On the test set, we achieve 66.3% MIOU on NVIDIA RTX 2060Super, which is 1.55% higher than the STDC net.