AI Classification
Running ONNX AI models on recorded seed images for classification
The AI Classification screen is where you run deep learning models on your recorded seed images. Select an ONNX model, configure duplicate removal and similarity search options, then execute the analysis on a single file or an entire batch of files. Results are written back to the HDF5 file and can be viewed on the Dashboard and Image Mosaic screens.
Prerequisites
An analysis file with recorded seed images (or a folder of files for batch mode).
An ONNX model (
.onnx) exported from Csmart Studio Desktop, with its accompanying JSON metadata file in the same directory.The Python runtime installed (from Runtime Environment).
Step-by-Step Walkthrough
1. Select the AI Model
Click Select Model to browse for your .onnx model file. After selection, the application reads the accompanying JSON metadata and displays:
Model Name: The name defined during export.
Model Description: A summary of the model's training dataset and architecture.
Model Classes: The classification categories the model can identify.
The selected model path is persisted in the Electron Store and restored on subsequent visits.
The ONNX file and its JSON metadata file must be in the same directory and share the same base name (e.g., model_v1.onnx and model_v1.json).
2. Choose Single or Batch Processing
Use the File or Batch Selector at the top to switch between modes:
Single File
Runs the model on the currently open analysis file
Batch (Folder)
Runs the model on all .hdf5 files in a selected folder
In batch mode, select a folder containing your analysis files. The application validates the folder and displays the number of files found.
3. Configure Duplicate Removal
The Duplicate Removal dropdown controls how the application handles near-duplicate seed images that may have been captured multiple times during recording:
Coffee - Default AI
Optimized for coffee seeds using AI-based similarity detection
Generic - Light
Conservative duplicate detection with fewer removals
Generic - Strict
Aggressive duplicate detection for maximum deduplication
Do Not Remove Duplicates
Skips duplicate detection entirely
For most coffee analysis workflows, Coffee - Default AI is recommended. It uses model-specific features to identify duplicates while preserving unique seeds.
4. Enable Similarity Search (Optional)
Toggle Similarity Search to enable feature-based similarity analysis during classification. When enabled, the model extracts embedding vectors for each seed, which can be used later for finding visually similar seeds across analyses.
5. Adjust Pixel Calibration (If Needed)
The screen displays the current Pixel/cm and Sieve Offset values stored in the analysis file. If these differ from your system calibration:
Use the increment/decrement controls to adjust the values.
Click Save to write the updated calibration to the HDF5 file.
Pixel calibration directly affects area and size measurements. Ensure it matches your hardware setup before running the analysis.
6. Run the Analysis
Click Run AI Analysis to begin. A progress bar modal appears showing:
Title: The current operation phase.
Progress Bar: Percentage of completion.
Message: Details about the current step (loading model, processing images, writing results).
Header: The file currently being processed (in batch mode).
The sidebar is locked during execution to prevent accidental navigation.
AI analysis can take several minutes depending on the number of seeds and model complexity. The application applies a 15-second timeout on individual IPC calls to detect and recover from stalled operations.
7. Review Results
When the analysis completes:
The progress modal closes automatically.
The total analysis counter is incremented.
The application navigates to the Dashboard screen to display results.
In batch mode, processing continues to the next file automatically.
Configuration Persistence
All settings on this screen are saved to the Electron Store:
Selected ONNX model path
Duplicate removal mode
Similarity search toggle
These are restored automatically when you return to the screen.
Troubleshooting
"Select Model" shows no files
Wrong folder selected or missing .onnx files
Navigate to the folder containing your exported ONNX model
Model description not showing
JSON metadata file missing or misnamed
Ensure the .json file is in the same directory as the .onnx file with the same base name
Analysis fails with timeout
Python runtime issue or corrupted file
Check runtime installation; try re-opening the analysis file
Progress stalls at 0%
Model loading failed or GPU out of memory
Check Runtime Environment for GPU status; try a smaller model
Batch mode finds no files
Selected folder contains no .hdf5 files
Verify the folder path and file extensions
Pixel calibration mismatch warning
File was recorded with different hardware settings
Update pixel/cm and sieve offset values before running analysis
Last updated