Printed label defect detection using twice gradient matching based on improved cosine similarity measure

Many vision-based methods for printed label defect detection have been proposed to replace inefficient manual inspection. However, due to the existence of artifacts and noise regions, it usually leads to a large number of misjudgments. Also, since most of the printed labels are non-rigid, they are prone to local deformation, which will cause lots of artifacts after image subtraction. This paper proposes a novel printed label defect detection framework (PLDD), which performs twice gradient matching based on improved cosine similarity measures. The overall idea is based on comparing a golden master (GM) image with test images, thus the GM image is demanded. Specifically, latent defect candidates will be extracted firstly from RGB sub-images for artifact elimination. Mask mechanism is also introduced to eliminate the influence of background gradient features around these defect candidates. Experiments compared with existing methods are conducted with three industrial datasets. The results exhibit that PLDD achieves a high mean F1 score (0.9702), and only 103 false positives (FP) occurred in 44,628 ground truths. Defects are being detected in real-time with an average time consuming of 0.26362 s.