RGB Images¶
Standard RGB (Red, Green, Blue) images are the most common data type supported by Phenoscape. This format is ideal for analyzing visible color patterns in biological organisms.
Data Format¶
- File Types:
.jpg,.jpeg,.png,.tiff - Channels: 3 (Red, Green, Blue)
Data Requirements¶
Image Quality¶
- Resolution: Minimum 224x224 pixels (higher resolution preferred).
- Focus: Sharp, well-focused images. The algorithm was tested on museum specimens but it should work on natural images as well.
- Lighting: Preferably consistent, even illumination.
- Background: Clean background or pre-segmented regions of interest.
Preprocessing Recommendations¶
- Segmentation: Remove background to focus on the organism
- Standardization: Consistent orientation and scale (preferably). This can be managed using proper augmentations.
- Quality Control: Remove blurry, overexposed, or underexposed images
Directory Structure¶
Organize your RGB images in the following structure:
Configuration¶
For RGB data, use these settings in your config.yaml:
data:
data_dir: "path/to/rgb/data"
output:
output_dir: "path/to/output/directory"
model:
name: "simclr"
backbone: "vit_l_16" # or resnet50
weights: "ViT_L_16_Weights" # or IMAGENET1K_V2 for resnet50
augmentations:
input_size: 224
Data Augmentations¶
RGB images support extensive augmentation options:
- Geometric: Rotation, flipping, cropping, perspective changes
- Color: Brightness, contrast, saturation, hue adjustments
- Noise: Gaussian blur, random erasing
- Advanced: Elastic deformation, thin plate spline
Example Usage¶
This will automatically train the model and generate embeddings (csv file) in the output directory. We also provide the option to generate embeddings from a pre-trained model:# Generate embeddings
python embeddings/generate_embeddings.py \
--model-path outputs/rgb_model.ckpt \
--data-dir data/rgb_images \
--output embeddings_rgb.csv
Related Topics¶
- Multispectral Data: Extended spectral information
- Training: Model training procedures
- Examples: Practical usage examples