Label defect detection algorithm based on twin network

Abstract

The invention discloses a label defect detection algorithm based on a twin network. The method comprises the steps of S1, obtaining of a training set and a test set; s2, network establishment and training; step S3: test set verification. According to the technical scheme, a twin network label defect detection system is built; the tag data set is input for training, then softmax is adopted for classification, only several types of tags need to be trained, defect detection can be carried out even if the type of the tag to be detected is not in the training set during testing, the workload of preparation work can be effectively reduced, the detection efficiency is improved, and the cost is reduced.

Description

technical field [0001] The invention relates to the technical field of label manufacturing, in particular to a label defect detection algorithm based on a twin network, which can be used to improve label management. Background technique [0002] As the carrier of product information, commodity labels contain a lot of information and play an important role in product management. However, there will be problems such as printing defects, missing numbers, missing numbers, broken numbers, and damages on commodity labels, which will have a huge impact on product management. There are hundreds of millions of labels in circulation in the market, and the quality of labels is of great importance. Before labels enter the market, it is extremely necessary to conduct quality inspections. [0003] At present, manual detection is mostly used for label detection, which obviously has defects such as low efficiency and low accuracy; there are quality inspection instruments in the market, whi...

AI Technical Summary

Problems solved by technology

[0003] At present, manual detection is mostly used for label detection, which obviously has defects such as low efficiency and low accuracy; there are quality inspection instruments in the market, which are bloated, expensive, difficult to maintain, and only support offline detection; the popular OCR solution pursues comprehensiveness Universality, efficiency and accuracy are inevitably not high

At present, the camera is used as an image acquisition device to identify barcodes, QR codes and characters to detect dynamic labels. The main schemes are template matching, background difference, and frequency domain analysis. However, template matching cannot deal with rotation and scaling problems. The detection label needs to be completely aligned with its correct label, which has high requirements on the network; while the background difference needs to match the samples of the background template with the training labels one by one. The training samples are more complicated. Like template matching, the correct label template of the label to be detected needs to exist In the training library; based on the frequency domain analysis method, there is a problem that false detection is easy to occur when the normal image and the defect frequency are similar

Images