Standard Operating Procedure

Operation of the Csmart-Digit Device

Purpose

To define the standardized procedure for operating the Csmart-Digit device to ensure consistent, accurate, and reliable analysis of green coffee samples.

Scope

This SOP applies to all qualified personnel involved in the operation and interpretation of results generated by the Csmart-Digit device. It does not cover AI model development, dataset generation, or hardware servicing.

Role	Responsibility
Operator	Execute analysis and perform basic troubleshooting
Analyst	Validate output, interpret data, and provide feedback
QA Supervisor	Monitor accuracy and compliance
Maintenance Tech	Maintain and calibrate device

Definitions

• AI Model: A deep neural network used by Csmart-Digit to classify seeds. Each model consists of two files stored in the same folder:

  • .onnx – contains the model weights and architecture.

  • .json – describes the model metadata, class labels, defect scores, and grading parameters.

Materials and Equipment Required

• Csmart-Digit device

• Csmart-Digit communication token

• Csmart-Digit software

• Laptop or desktop computer (minimum specs) with the software installed

• Tray, funnel, and brush

• 300g sample of green coffee beans

Safety Conditions

• Avoid physical interference with the belt conveyor and any moving parts while the device is in operation.

• Place the device on a dedicated, stable, and level surface capable of supporting its weight.

• Do not operate the device with wet hands or in humid environments.

Operating Temperature	15–28 °C
Maximum Relative Humidity	75%

Startup

• Power on the device, according to Powering On Procedures.

• Power the computer and launch the Csmart-Digit software.

• Ensure all cables and connections (USB from Token, USB from camera, and power cable) are correctly in place.

Mechanical and Electrical System Check

Before beginning any analysis, ensure that all mechanical and electrical systems are properly functioning.

Open the Acquisition Settings page in the Csmart-Digit software.

Perform the following activations:

• Click Camera On to initiate the vision system.

• Switch Light to On to verify illumination.

• Switch Belt to On to start the movement.

Confirm that the following configuration parameters are set appropriately:

• Pixel/cm ratio (used for accurate sizing and screen classification).

• Min and Max Area filters (ensure these values are aligned with the size range of the seeds in your sample).

Belt movement and illumination:

• The belt should run smoothly.

• Illumination should be uniform and clear across the entire field of view.

Note: Light flickering is perceived due to the use of overdrive.

Attention: Ensure there is no direct light source or sunlight hitting the capture area

Mechanical Cleanliness:

• Ensure the hopper is empty and contains no residual seeds from previous analyses.

• Inspect the belt conveyor to confirm it is free from:

  • Leftover seeds

  • Dust

  • Foreign matter

• If cleaning is needed:

  • Use a soft brush to gently remove any debris from the hopper and visible parts of the conveyor.

  • If required, clean the belt using a lightly moistened disposable wipe:

    • Hold the wipe gently at the end of the moving belt.

    • Apply light pressure to clean the surface.

    • Do not over-moisten; the belt must remain dry to ensure proper imaging and transport.

Caution: Never apply water, alcohol, or other cleaners directly to any part of the machine. Excessive moisture can damage optical and electronic components.

Sample Preparation

• Weigh 300 grams of green coffee beans.

• Ensure the sample is statistically representative of the coffee lot it aims to represent.

• Place a collection tray at the end of the belt conveyor. The tray must be capable of collecting at least 300 grams of coffee.

• Ensure the sample is within the nominal seed range supported by the device:

Nominal Seed Range	2 - 20mm
Sample Standard Weight	300g

Capturing the Sample

• Pour the 300g sample into the hopper compartment.

• Open the New Analysis page in the Csmart-Digit software.

• Name the analysis with a descriptive name (e.g., sample 12345) and click the New Analysis button.

• Click Camera On to start capturing.

• Switch Light On to start the lighting system.

• Switch Belt On to start the horizontal transportation of the seeds.

• Observe the seeds flowing and passing through the vision system

• Every seed is marked on the screen with a red dot, indicating it was detected and properly stored.

• Allow the machine to fully process the sample (approx. 2 minutes).

• Ensure no seeds remain unprocessed in the system.

• Click Stop to finalize the capture phase.

Running the AI Model

Upon completion of the capture, images are saved to a .hdf5 file, and the system automatically displays the Artificial Intelligence page.Choose the AI model for the newly captured sample. Verify that its attributes and training dataset are consistent with the sample being analyzed.

Tip: An AI model can detect patterns that may be challenging for even a trained human eye to recognize. However, it is less capable than humans at making generalizations and correlations. Using a model on a coffee type it was not trained for will likely produce poor results, even if it performs well for its intended coffee type.

Model files are user-managed, and no standard storage location is defined.

After selecting the model, choose whether to run a pre-stored model to detect and remove connected seeds. The model options are:

Model Name	Function	When to Use
Coffee – Default AI	Standard model optimized for coffee seeds	Default choice for coffee seed analysis
Generic – Light	Applies a light watershed algorithm for detecting and removing connected seeds.	For roasted coffee or unknown seeds
Generic – Strict	Applies a stricter watershed algorithm for more aggressive separation of connected seeds	For roasted coffee or unknown seeds
Do Not Remove Duplicates	Skips duplicate removal, keeping all images	When the analysis has already been run with one of the models above

When classifying coffee, use the Coffee - Default AI model.

The final parameter to configure is Similarity Search, described in the table below:

Tool Name	Function	When to Use
Similarity Search	Removes subsequent duplicate images caused by performance errors in the object tracking algorithm	When repeated images are visible in the Image Mosaic

Tip: If the AI analysis is currently running on the CPU, refer to the GPU Installation Guide to take advantage of your graphics card’s processing power for faster AI network execution.

Before Reviewing the Results

After the AI finishes processing, the software automatically navigates to the Dashboard page, where the analysis results are displayed. Before reviewing the results, take note of the following key features:

Setting	Location in Interface	Description
Relative to Weight / Relative to Count	Top right of the Dashboard screen	Allows the data to be displayed either relative to the total weight of the sample or relative to the total count (occurrences) of individual seeds.
Ok / Defects / Both	Bottom of Screen Size Distribution panel	Control how the screen size distribution is presented: only seeds free of defects (OK), only defective seeds (Defects), or both combined (Both).
Total-relative / Group-relative	Bottom of Screen Size Distribution panel	Determines the denominator used in percentage calculations when OK or Defects is selected. Total-relative uses the sum of OK + Defects as the total, while Group-relative considers only the sum within the selected group (OK or Defects).

Notice: These features only need to be checked once, as their settings are saved in the software and will persist for future sessions, using the last selected configuration.

Common Occurrences

Use the table below to identify common occurrences, understand their possible causes, and follow the recommended actions.

Occurrence	Possible Cause	Recommended Action
Low Sample Weight Warning	Input weight is below the minimum recommended weight	Re-scan the analysis with a larger sample, or use the current results considering a higher margin of error
Density Data Missing Warning	AI model does not contain a Density Map, or the Density Map is incomplete	Review and update the Density Map values in the model
Estimated weight is significantly different from original weight	Density Map does not match the actual sample characteristics	Review and update the Density Map values in the model
Only a fraction of seeds detected	Acquisition issues causing CPU bottlenecks, skipped frames, or lost data	Re-scan the analysis under stable processing conditions; check CPU/GPU load during capture. Close any unnecessary background processes and other software before re-scanning
Equivalent defects error range is too high	Input weight is below the minimum recommended weight, resulting in a higher error range	Re-scan the analysis with a larger sample, or use the current results while accounting for a higher margin of error
Type/Grade value is outside the expected range	Input weight is below the minimum recommended weight, or the model used in the analysis is not compatible with the analyzed coffee	Check the entropy value and consider using a different model
Entropy is too high	The model used in the analysis is not compatible with the analyzed coffee	Consider using a different model or adjusting the existing model to improve generalization
Inference Confidence Level is not adequate	The model used in the analysis is not compatible with the analyzed coffee	Consider using a different model or adjusting the existing model to improve generalization
Screen Size Distribution is outside the expected results	Scale conversion is incorrect, or the selected display values are not appropriate	Check the pixel/cm setting for the analysis file and review the OK / Defects / Both and Total-relative / Group-relative options