Artificial intelligence is revolutionizing the verification of variable data in the food industry. Even Chinese characters on a label can be verified with one click: without any knowledge of Chinese or programming. A sophisticated optical character recognition (OCR) algorithm makes this possible.
Companies that fail to mark products or their packaging correctly risk not only a loss of image but also penalties, product recalls or contractual penalties from retailers. For this reason, checking these markings is essential. Optical checks are often automated, but in the case of plain text, additional expensive and error-prone human inspection is usually required. Here, OCR can now provide a remedy.
Industrial image processing and variable data
Optical technologies developed in the field of industrial image processing are of particular importance for labeling inspections. While the possibilities for inspecting machine-readable code are already well advanced, the automatic inspection of plain text with cameras has been subject to considerable limitations to date, as the OCR (Optical Character Recognition) algorithms used to read the recorded markings are highly error-prone. A reliable reading process could only be ensured if the variable data was applied in the best possible way.
In packaging processes in the food industry, however, it is common for labels to be printed very quickly and therefore perfect print quality is by no means guaranteed. The error rates during reading are correspondingly high, and time-consuming manual random checks are still widespread.
Another disadvantage of solutions used today is the high effort required for adaptation to the respective conditions. Smart solutions, such as those used for checking machine-readable code, have not yet been able to establish themselves because of the limitations described above.
Technical breakthrough in reading plain text
We are currently witnessing a technical breakthrough that will enable the automatic reading of plain text even in the food industry with all its complex requirements – such as fast cycle times, fluctuating font quality and harsh environmental conditions. Neural networks, a specialty of artificial intelligence generated using Deep Learning methods, can learn to recognize plain text very reliably even under difficult conditions: on irregular backgrounds, with poor print quality or unusual fonts.
Neural Networks and Deep Learning
Neural networks which are created with Deep Learning methods imitate the functioning of the human brain. Artificial neurons serve as a storehouse of information and exchange it with each other via connections between the neurons of individual layers and regions, following the example of biological synapses. During the intensive learning phase, the neurons receive feedback on the ways in which they have contributed to a current result. Over time, based on this feedback, they change and the connections between neurons are also subject to a constant process of development. This optimizes the overall neural network's ability to perform classification or approximation tasks.
Revolution through independent learning
For digital image processing, this represents an enormous step forward. Instead of having to program the system in manual and detailed precision work to determine which features it must look out for in order to clearly recognize an object in a camera image – for example, a letter or a number – the algorithm can learn for itself what the relevant features are. The human brain does a masterful job of such recognition, even when characters deviate from the norm. But we can rarely explain exactly why we know which letter is represented. With Deep Learning, we no longer have to explain to the algorithm by which means this is to be recognized; instead, it receives feedback during the learning phase whether it has recognized imprinted letters and numbers correctly or whether the extraction was faulty. In this way, it learns better and better to recognize what the specific criteria are that distinguish the individual characters.
Although neural networks have been the subject of research for many years, they are only gradually gaining ground in industrial applications. The appropriate computer hardware capable of simulating very large and complex networks has only recently become available at an affordable price.
What advantage do these new processes offer for practical use?
- Robust, reliable and fast reading even where quality varies greatly
- No adaptive programming work when booting a new installation
- No adjustments when the line configuration changes
- No manual rechecking
The complete Safe-Ident OCR system from Strelen Control Systems GmbH
Strelen Control Systems GmbH has developed a system for character verification that can be quickly and easily integrated directly into production lines. The camera and lighting are installed inline – even in the smallest of spaces – and the computer with the sophisticated Deep Learning software is located in a dust- and moisture-proof stainless steel cabinet. The system is operated via a touchscreen with an intuitive user interface.
Reliably verifying Chinese characters with OCR
Food exports to China are subject to strict regulations. Upon import, Chinese authorities strictly control whether all information is correctly noted on the labels – product name, best before date, place of production, etc. If the information is missing or incorrect, the entire shipment can be rejected – a disaster for any producer. Therefore, it is essential to ensure that correct and proper labels are applied during production. Often, labels are checked manually by employees before production starts, but they are usually not able to check the correctness of the applied Chinese characters.
Digital image processing can provide a remedy here. A scanning station takes the labels in question and checks all the desired graphic elements – such as logos – any codes (Datamatrix, EAN, barcode, etc.) and printed text for presence and correctness. For the latter, the highly developed optical character recognition (OCR) algorithms are used. The system is quickly and easily taught the correct data, which is then used for subsequent verification.
The learning process
For teaching, the reference label is recorded with the camera station. The user-friendly software displays the image and after the label has been selected, the relevant elements are highlighted in the menu and outlined with the mouse – for example, the barcode, the logo, various data (variable and static) and information in Chinese script – for example, the product name, the destination country and the place of manufacture. With a click on “Save” the data is stored in the system.
The examination
Before production, a label can now be randomly placed in the scanning station and inspection can start. By entering the article number and the variable data – for example, production date, best before date or batch number – the system can compare the inserted label with the stored reference and check whether all data is present, correct, and legible. The variable data can alternatively be retrieved from a database. If all fields to be checked are correct, production can start, and it is ensured that the correct labels are present!
Quality is key
However, correct, error-free and legible labels are not only of great importance when exporting to China. Providing high-quality and safe products is a core concern of every food manufacturer. Correct and legible data on product labels is essential in order to meet quality standards. Compliance with a wide range of laws and standards goes a long way toward protecting brand reputation. In addition, correct data helps reduce the risk of a product recall, with all the cost, waste, and image damage that entails. Quality requirements for retailers have increased dramatically in recent years, recall rates have increased and penalties have soared.
