09/14/2017

Case study: Real-time food inspection using deep learning

Jacob Zweig
Co-Founder, Principal Data Scientist

In a recent project, Strong worked with Sunsweet Growers, a grower-owned prune cooperative and the world's largest prune distributor. For over 100 years, they have been producing and delivering prunes to customers around the world. In this project, we worked to automate their inspection process — a painstaking and tradition-rich process established near the outset of the company — using computer vision and deep learning.

Building on tradition with the state-of-the-art

Since its founding, Sunsweet had perfected a manual process that ensured the product going out the door was of the highest quality. Although painstaking, this process had earned the trust of stakeholders — something this new solution would have to earn as well.

In applying state-of-the-art machine learning to this problem for the first time, we had to be mindful of important external considerations, including the fact that it would ultimately be deployed as an edge device in locations with poor remote connectivity.

Our solution

Strong Analytics designed, built, and deployed a deep-learning based computer vision solution built on Strong Vision to automatically detect defects in real-time from distributed food collection and grading sites throughout the world.

There were three major components to the solution:

  • First, there was the hardware device that collected images of each prune from multiple cameras in a portable form factor to allow on-site automated inspection grading. 
  • Second, these data were streamed to a cloud-based computer vision pipeline that built holistic visual representations of each prune and then analyzed them using custom-trained deep learning models to identify and classify visible defects. 
  • Third, data from each image, prune, and batch were summarized and streamed to a data warehouse for further analysis and visualization.

Critically, to ensure that we were as accurate as possible in defect analysis, our software enabled continuous 'human-in-the-loop' annotation of new data. Annotators could inspect any batch and provide feedback to the models about the presence or category of any defects, which the model ultimately used to continue training and improve accuracy.

Harold Upton, CTO of Sunsweet, shared this about the solution:

Strong designed, built and deployed a solution that integrated computer vision models with a monitoring/reporting dashboard that our team relies on throughout each production season. Since its initial deployment, as new challenges and opportunities have arisen, Strong remains a valued collaborative partner to Sunsweet.

Each season, this system processes millions of prunes to ensure that the highest-quality products are delivered to Sunsweet's customers.

Share: