Independent Stevedoring Ltd

Log Counting utilising Machine Learning

Background

Independent Stevedoring Limited (ISL) is a stevedoring company specialising in project cargoes, bulk cargo, fresh produce and added value commodities. As a specialist in the industry, ISL has an international reputation for the skilful handling of a variety of mixed break bulk cargoes.

THE CHALLENGE

Cucumber developed a new mobile logistics solution to replace ISL’s legacy system used to validate, track and load logs onto ships.

ISL wanted the ability to automatically count the number of logs in each load to validate the manual count of logs scanned.

The sample photo of logs below shows the environment where successful counting had to be achieved. There is shadow covering some of the logs as well as logs from another load being visible to the right of the photo.

OUR APPROACH

The existing system was using mobile devices to scan the barcodes on logs as they are being loaded onto ships. By using the camera to take a photo of the logs, could a computer program count the logs accurately enough?

Cucumber implemented an image recognition engine using a computer vision for a machine learning model. Powered by AI (artificial intelligence), the image recognition engine enables computer assisted identification and counting of logs.

The issues

The image recognition engine had to be taught what was the end of a log and what wasn’t. But the logs are loaded 24/7 in all weather conditions which meant the image recognition had to be accurate in

Bright sunlight
Rain
At night with artificial (and poor) lighting
Different angles

The image recognition engine also had to not count any of the loading equipment as a log as well as cope with the log bundles being uneven so that some logs were further away from the camera than others (see example photo earlier in this article).

Training log recognition

In order to improve recognition performance, the results were judged by a human against a range of photos in all conditions and fed back into the engine. In this way, the engine could be taught by being “told” when it had identified a log that wasn’t a log and when it had not identified a log.

THE OUTCOME

After the image recognition engine was built and incorporated into the live system, supervisors would be notified when the automated count of logs differed from the manual count on the wharf. The supervisor could then correct the final image by adding and removing logs.

Cucumber built reporting on the success of all photos so that a percentage of logs counted could be generated per ship.

We have refined the engine further when it was identified that there was a particular issue with certain wharf conditions. We were able to use the real photos taken and “teach” the engine what was wrong.

The addition of the image recognition engine provides ISL with increased accuracy of log counts while lowering the operational load of the stevedores, allowing them to focus on other tasks. This results in increased throughput and efficiency when loading ships. The photos generated by the system are valuable for auditing and traceability purposes as an additional benefit.

By continuous improvement, log counting accuracy for the last 2 years of operation has increased from 94.0% to 96.4%.