1. Context and Users

Understanding App Users

The main users of the app will be commercial rice value-chain actors in Ghana, including buyers of paddy rice at farm gate or mill level, and buyers of milled rice at rice mills. The application will be used to support rapid buy/reject or price-adjustment decisions based on visible quality indicators.

The app will be in English only and designed for literate or semi-literate SME users.

User Devices

The MVP should run on devices commonly used in Ghana, such as Samsung Galaxy A series, Tecno, and Huawei phones. The application must run fully offline for image analysis. GPS capture is optional (user opt-in), while date/time must be stored. The app should be optimised for Android 9 or higher, and low-to-mid range hardware.

2. Offline Workflow and Required Functionality

What the app must do

• Create a user profile, capturing name/username, role, and organisation (optional).
• Take a photo of a rice sample (or select one from their phone)
• Run the provided AI model on that image
• See rice quality results instantly in a clear, understandable format

Core screens

Home, Capture, Results, History, Export, and App Information. Results can be shown as a simple summary first, with detailed metrics accessible on demand. The app must display the full set of model outputs from the UNIDO AfricaRice Quality Assessment Challenge.

Overall User experience requirements:

The app should:

• Guide the user when taking a photo
• Show results in simple language, not just raw numbers
• Return results in a few seconds
• Work reliably on a lower-spec Android phone

3. Input Data and Capture Guidance

Image Capture Guidance

• Blue background: The app should guide users to capture images on a blue background, consistent with the training dataset.
• Basic validation: The app should provide on-screen guidance to help users spread grains in a single layer and avoid shadows or blur. Image capture will not be overly constrained, reflecting real field conditions. Basic validation should warn users if an image is too dark, blurry, or unsuitable for analysis.

4. Model Outputs and How Results

Displaying Rice Quality Metrics

The app will report:

• Grain count and structure (total, broken, long, medium)
• Grain colour composition (black, chalky, red, yellow, green)
• Kernel shape (average length, width, and length-to-width ratio)
• Colour profile using CIELAB values (L, a, b)

and then apply industry classifications provided by AfricaRice to display conclusions in plain language indicating grain shape, grain length class, chalkiness status, consistent/mixed status and milling return grade.

Raw CIELAB (L*, a*, b*) values will be displayed, with interpretive thresholds to be added in future iterations.

Below are some definitions for polished rice:

• Milling grade classification (Less than 5% broken = premium, 10% broken = grade 1, 15% broken = grade 2,  and 20% broken = grade 3)
• Grain shape classification (Length Width Ratio (LWR) <2.1=Bold; LWR 2.2-2.9=Medium; LWR>3 Slender)
• Long percent > 90% = Long grain, Medium percent > 90% = Medium grain, Short percent > 90% = Short. If none of the above is true, it means MIXED (for example, Long percent = 30; Medium percent = 40 and Short percent = 30, then MIXED).
• Chalky grain classification (<20%=not chalky; > 20%=chalky)
• Black Percent > 10%: damaged or defective grains
• Green percent > 10%: immature grains
• Red percent > 10%: grains red strips
• Yellow percent > 10% in non-parboiled polish rice: fermented grains

These values must be computed by running the supplied model - not hard-coded.

The Top 3 solutions from the UNIDO AfricaRice Quality Assessment Challenge are available here. We recommend using the 3rd place solution as your starting point, as it is the most lightweight and mobile-friendly of the three. You are welcome to adapt, simplify, compress, or optimise any of the provided solutions to improve performance, reduce size, or enhance suitability for offline mobile deployment. The goal is to ensure your final implementation is efficient, practical, and ready for real-world field use.

5. Data Storage and Export

• All data will be stored locally on the device.
• All user data (from sign up) and image processing data (time, location, metrics etc) stored in a database.
• The app will maintain a limited scan history (e.g. last 100 scans).
• Results can be exported as a CSV file via email when connectivity is available.

6. Technical Requirements

• The app must use one or more of the three winning models from the UNIDO AfricaRice Quality Assessment Challenge
• The model must run on-device (e.g. using TensorFlow Lite or ONNX)
• The app must load the model locally, perform inference on-device, and display the model version within the application for traceability.
• The same model and preprocessing pipeline must be used for all test images
• The app must run without internet access
• The app must remain stable when used offline and provide clear, user-friendly error messages when analysis cannot be completed.

7. Legal

The app must display a disclaimer statement when users sign up and load the application. The disclaimer must state that the tool is intended for indicative, field-level quality assessment and does not replace laboratory analysis or provide food safety certification.

Intellectual property for the solution will be co-owned by UNIDO and AfricaRice.