Identify Direct / Machine Vision Technology (MVT) have developed a system using an early example of AI to solve the classic problem of mis-labelling of food cans (tins). Mis-labelling is not only bad for the customer e.g. food allergies, vegetarian / meat etc but historically has also been very costly to the manufacturer due to massive fines imposed by supermarkets for product recalls.

Historically

A large canning company in the UK, with 26 production lines over two sites, was particularly susceptible to mis-labelling due the a very large number of products for customers, which include several major supermarkets. Obviously, our customer was keen to avoid any product recalls due to mis-labelling and the resulting fines.

MVT developed a high speed can reader, which used OCR (Optical Character Recognition) to read the inkjet can code printed with the Best Before Date on the end of the can and coupled this with a laser barcode scanner to read the linear (1D) barcode on the label. If the two matched, all was good. If there is a mismatch, then the system automatically stopped the line. If the codes cannot be read, then the individual can is rejected for manual sorting and re-cycling.

This OCR was an early example of AI and used a neural network approach to train the font on multiple cans off multiple printers on multiple lines. This use of AI allowed the system to adapt to real world variables with the very low resolution dot matrix font (the characters are only 7 dots tall and 5 dots wide). Characters are not continuous (because they are made up of dots but yet neighbouring characters can touch or overlap!) These systems have remained operational to the present day preventing all mis-labelling and the corresponding fines.

A New Problem 

 Many supermarkets are now introducing DataMatrix (2D barcodes) to their labels. These cannot be read with the traditional laser scanner but require an imager-based reader. These pose a new challenge to the system integrator, in that ‘motion-blur’ becomes significant, especially on canning lines which typically run at 2 metres per second and up to 13 cans per second.

To make matters worse, the manufacturer was told by the supermarket that they needed to place a 2D barcode in the overlap region of the label, for use in ‘multi-pack’ cans. The reason being that they did not want any unscrupulous retailer splitting a pack of cans and selling them separately. Therefore, the barcode cannot be visible on the finished product to prevent it being scanned at the checkout.

Key factors to finding a solution were:

  • 2D barcode had to be read after the leading edge of the label is glued and before the trailing edge is fully wrapped round (it then overlaps the leading edge and covers up the barcode).
  • Can speeds up to 13 cans per second, causing blurring of the image due to motion.
  • Very bright illumination needed to cope with the very short shutter speed which is needed to counteract the motion blur.
  • Shiny labels causing stray reflections from very bright direct illumination.
  • Each time the glue gun fires it causes strands of glue which land on neighbouring equipment and would obscure a camera lens within minutes.
  • A very narrow time window and a very narrow angle in which the barcode on the partially applied label is actually visible.
  • Can spinning as the label is applied needing extremely accurate triggering of the image capture

The Solution

Identify Direct / Machine Vision Technology developed a solution using a Keyence SR-X300 camera-based reader.

  • A position was chosen in the rotation of the can being labelled where the barcode was square on to the reader and visible, and a very precise trigger sensor was chosen to reliably trigger the reader at this instant in time.
  • A long standoff distance was chosen (a feature of the SR-X300) to enable the reader to be well clear of the glue.
  • The strobe type illumination was set to maximum to compensate for the very short exposure time and large standoff distance.
  • The SR-X300 has a feature where it can use polarised light and ‘cross polarisation on the lens’ to virtually eliminate all specular reflections from the direct illumination. This was a key feature.
  • The AI feature was used which, together with capturing 3 successive images, greatly increased the reliability of the reading; (if the first image could not be read then the second was tried and if necessary, the third).
  • A PC was used to display every image in real time for the sake of the operators. This enables them to very quickly work out what is needed to be done when the successful read rate decreases; for example clean the lens or re-align the camera after a product change.

The system is in production, 3 shifts per day and 6 days a week and has outperformed expectations. It has not required any specialist intervention from us the system integrators after the initial setup and training of the customer.