In postal and parcel processes, mailing addresses have to be extracted from images. For humans, not a problem, because we can quickly recognize patterns and select, understand, and evaluate specific information from a large quantity of data. However, teaching this ability to machines has required breaking processes down into pieces. But with a new era of AI based on deep learning, training machines has become much simpler.
Teaching robots how to singulate
A case in point is teaching robots to singulate individual pieces from a bulk stream predominantly consisting of e-commerce parcels. They are often not only small, but also oddly shaped and sometimes poorly packaged – and they therefore tend to easily shift, roll or tilt when processed on belt systems. For a robot to tackle the problem of picking from a bulk stream, it has to be taught to “see” the location of individual parcels. Here Körber Supply Chain uses a system based on deep learning that helps to accurately identify the objects and their shape with a classification scheme. The recognition system also identifies the optimal gripping point for the robot to ensure that the parcel is transported safely. The algorithm at the heart of the classification scheme can learn from each new image and fine-tune its assessment criteria. Image processing that is enhanced by artificial intelligence can create a fully integrated and automated solution.
Reading addresses on small parcels
Another example of machine learning is the automatic reading technology optimized especially for small parcels.
The key success factors are:
- the reliable detection of the receiver address block based on pre-trained label types,
- the correct address interpretation, even if the address is syntactically incorrect or elements are missing, and
- the accurate segmentation of single characters, especially for small fonts that tend to blend together.
The solution from Körber Supply Chain also extends to the reliable recognition of handwritten addresses.
Fingerprint for tracking
Additionally, machine learning has helped to enhance the field-proven Fingerprint technology. Initially developed to track flats in sorting machines without having to apply barcodes or labels, today enhanced Fingerprint technology is used to track parcels. Only one side of a parcel is sufficient for reliable recognition, even if the address label is not facing the camera. Because any side of a parcel can be used for tracking, requirements concerning camera hardware are lower.
Deep learning to comprehend complicated patterns
What most of these examples have in common is that they employ deep learning, a machine learning technique that teaches computers to do what comes naturally to humans: learn by example. Deep learning combines advances in computing power and special types of neuronal networks to learn complicated patterns from large amounts of data. In the case of reliably recognizing handwritten addresses, it involves recording and labeling a data set for a specific country’s alphabet and teaching a pretrained neuronal network with the new dataset.
