CASE STUDIES

Cases in operation and in validation

Where OpenScience’s platforms already work, and what is still being validated.

Automated coffee bean classification

Problem. Manual defect classification, subject to variation between graders and with no per-sample record.

Solution. A two-camera system (NIR and VIS), per-bean segmentation and the OS4CI classification engine, with multiclass boosting models, integrated into the receiving workflow.

Result. Standardized classification, a historical per-sample database and mass estimation by sieve size.

VIS image of coffee beans and classification map generated by the Imajan HS4017

Multispectral system for rice classification

Problem. Rice classification dependent on visual inspection, with low traceability and long per-sample time.

Solution. Multispectral equipment with a conveyor, sequential illumination and a 20-megapixel camera, delivered to an official laboratory, with software that measures the size and mass of each grain and separates the long slender, long, medium and short classes.

Result. Four classes by size, with size and mass per grain and a per-sample report with image records. Defect analysis is under validation. The image alongside shows the mosaic of valid grains in the operator panel.

Operator panel of the rice classification system, Grains and Mosaic tab, showing the valid grains and their dimensions

Foreign material detection on the conveyor

Problem. Contaminants (plastic, paper, string) mixed into tobacco mid-process, caught late and at a rework cost.

Solution. An industrial camera with dedicated illumination on two reflectors and variable selection via a GA-GRASP optimizer to separate tobacco from foreign material in process time.

Result. A proof of concept with a real-time monitoring panel and configurable alarm criteria.

Tobacco with foreign materials in false-RGB, 620, 520 and 460 nm — Drag the bar to overlay the pixel-by-pixel classification (Random Forest, 5 classes) on the false-RGB image (620, 520 and 460 nm).
Tobacco Metallized Metal Foam Plastic

Pixel-by-pixel Random Forest classification, five classes — Drag the bar to overlay the pixel-by-pixel classification (Random Forest, 5 classes) on the false-RGB image (620, 520 and 460 nm).
Tobacco Metallized Metal Foam Plastic

Soybean grain defects

Problem. Defect classification (burnt, moldy, greenish) with variation between graders.

Solution. Classification models using morphological and spectral information.

Soybeans in false-RGB, 640, 520 and 410 nm — Drag the bar to overlay the classification map on the false-RGB image (640, 520 and 410 nm).

Classification map of the same beans, each color is a class — Drag the bar to overlay the classification map on the false-RGB image (640, 520 and 410 nm).

Sensory analysis 5.0: objective and traceable

To deliver objective sensory assessments and recognize commercial coffee blends at an operating cost of R$10.00 per sample (variable costs), we use high-resolution mass spectrometry (a Waters Xevo G2-XS QTOF).

Sensory coffee analysis by mass spectrometry

Real-time beer analysis

Problem. Evaluating beer by conventional methods requires sample preparation and analysis time.

Solution. Real-time chemical fingerprinting: thousands of molecular features acquired in 2 seconds per run, with no sample preparation.

Result. Process already applied to more than 30 beers.

Real-time beer analysis by chemical fingerprinting

Tomato defects by computer vision

Problem. Manual defect classification, subjective and with low repeatability between graders.

Solution. Imaging and classification models trained and validated on independent samples.

Result. 91% accuracy in defect classification, in OpenScience in-house research.

Want to test it with your own samples?

Remote or in-person demo in Campinas, Brazil. The software can be evaluated with your own data.

REQUEST A DEMO