Custom Computer Vision for Livestock Health Screening
Developed a specialized image-model pipeline for assessing visual anemia-risk patterns from lower-eyelid conjunctiva images.
- Stage
- Research demonstration
- Industry
- Agriculture and Livestock
Summary
In brief
OB360 explored a custom computer vision pipeline to support visual screening of anemia-risk patterns in goats from images of the lower-eyelid conjunctiva. It combines synthetic data generation (DDPM), SAM-assisted annotation, Attention U-Net segmentation, hybrid deep-learning classification and explainable AI.
It is a screening-research capability, not a substitute for veterinary diagnosis.
This is a research and engineering demonstration. It is not a veterinary diagnostic, and no field deployment or diagnostic accuracy is claimed.
Challenge
The business challenge
Goat farmers need a practical way to monitor animal health regularly without transporting every animal for veterinary assessment. One visible indicator associated with anemia risk is the color of the lower-eyelid conjunctiva.
The relevant information occupies a small anatomical region, farm-captured images vary widely in lighting, angle and background, and labeled training data is limited. A generic classifier working on the whole image could learn irrelevant correlations.
Solution
The engineering pipeline
- Synthetic image generation with denoising diffusion probabilistic models (DDPM) to widen limited training data
- Mask preparation assisted by the Segment Anything Model (SAM)
- Attention U-Net segmentation to isolate the conjunctival region of interest
- Hybrid deep-learning classification of visual severity patterns
- Explainability with Integrated Gradients, LIME and Grad-CAM to check where the model is looking
- An integration direction for a simple farmer-facing application
The intended farmer workflow
-
Capture
The farmer photographs the goat’s lower eyelid with a smartphone.
-
Segment
The model isolates the conjunctival region from the eye, fur and background.
-
Classify
Visual characteristics are analyzed and a severity pattern is classified.
-
Decide
The farmer uses the screening result as one more signal: keep monitoring, or prioritize a veterinary visit.
Outcome
What the research demonstrates
The work demonstrates how specialized image models can be engineered around a real domain problem when general-purpose image services are not enough.
The same approach applies to other specialized image-analysis problems, in agriculture and in fields such as manufacturing quality, infrastructure and equipment monitoring.
Capabilities
Capabilities demonstrated
- Computer vision
- DDPM
- SAM
- Attention U-Net
- Hybrid deep learning
- Explainable AI
Related services
Discuss a similar initiative
Tell us about your business workflow, users and existing systems. We will discuss the architecture, engineering support and next steps that fit your needs.