Client situation
We were approached by an existing client, a dominant international postal and logistics organization, who was increasingly frustrated with failed Barcode Reading. The successful read-rate was almost 95%, but the 5% fail-rate was having a significant effect on parcel volumes.
For every failed barcode reading, manual intervention was required to avoid any potential bottlenecks. Consecutive unreadable labels had a knock-on effect that resulted in excessive delays contributing to sometimes missed deadlines and possibly broken contractual agreements. Of particular concern was the fact that the operators whose job it was to manually process these parcels, were under constant pressure to decipher the failed label details as quickly as possible to avoid the build-up of these costly problems.
Challenge: reduce failed reads
One considerable problem lay in that industry-wide systems were designed without an automated response to failure. Simply put, these sortation engines didn’t attempt re-reads, leaving manual intervention the only option to keep the parcels flowing.
A failed read can have many causes but in this case, it came down to two main issues: the poor label standards and, ironically, the parcels themselves. By nature, parcels come in all shapes and sizes, often with surfaces preventing the label being presented flat to the scanner. Also, parcels are often wrapped in plastic, rendering the label (partly) invisible.
Concerning the labels, they often don’t meet industry specifications and are in consistent in quality; sometimes lines are blurred, uneven, incomplete or damaged. The quiet (empty?) zones can be too small, or the bar ratios are infrequent (inconsistent?). Also, low contrast, image blur and poor printing can lead to binarization issues, literally leaving the barcode readers unable to tell black from white.
Any of these problems can produce a failed read. And every time a human has to get involved, it becomes costly. A 5% failure rate means one-in-twenty parcels are manually processed. With millions of deliveries sent every day, that’s a huge amount of wasted time and money.
How we solved it
The R&D team at Prime Vision needed to develop a system that could read any label, no matter the circumstances. The project was led by senior researcher Sjaak Koomen. “The key to the solution was this: Instead of re-designing the bar-coding equipment, we changed the way the components interact and, crucially, how they ‘think’. We built our solution on the assumption that all printed barcodes do not comply with industry specifications. We based our software on the barcode material experienced in real life. By utilizing a neural network and applying machine intelligence, we created a solution that can learn to correlate a deformed signal to an ideal signal. We’re considerably proud of this innovation.”
A particularly impressive innovation was the handling of blurred images. In an intricate process, the R&D team set about creating artificial black-and-white barcode signals and then manipulating them to simulate every possible distorted signal that could come back from a distressed label. Once paired, these signals are used to create a neural network, enabling machine intelligence to learn to connect a deformed signal to an ideal signal.
Simply said: when the system is presented with a disturbed signal from a parcel, it says to itself. “If I can see disturbance X, so I know this must be this sharp, crisp, ideal Signal”. The result effectively eliminates blur for any barcode.
Benefits of Barcode Vision
Researcher Sjaak: “Overall, the new solution can read 70% of the previously failed reads, taking the successful read-rate from 95% to 98.5%. ”To put this in perspective, if a postal company sorts 1 million parcels a day, 35,000 fewer parcels need to be sorted by hand than before our solution was introduced.”
Even by spending just 30 seconds to manually process a failed read, Prime Vision’s solution saves 291 working hours per day to process the 35,000 failed reads.
There are benefits for the equipment operators as well. The higher rate of successful reads considerably reduced the pressure on them. The reduced failure rate leaves over 300% more time to process the remaining failed reads at the same operating costs, reducing pressure-induced stress and secondary errors from hurried operators.
Conclusion
With the project fully operational and the client enjoying the benefits, Sjaak considers the project a success. “It’s remarkable that all these positive effects come from one solution, consisting of small innovations. It’s exactly these small steps that made the real difference, because together they add up to the 3.5% read gain. It’s this 3.5% that counts: even such a small percentage can turn loss into profit.”
