This will go through the objects, in pseudo-chronological order
Image Analysis
Zero-shot Learning
Pre-trained Models
CLIP (Contrastive Language-Image Pre-Training) by OpenAI
DINO (self-Distillation with NO labels) by Meta AI
Active Learning
Transfer Learning
Identification
Segmentation
characterization (All as Time-series Features)
opacity (Inside and outside the well)
Circularity
Compactness
Speed of formation
Size
Smoothness
Amalgamation (of characteristics)
Classification
2d colour function showing the pixel intensity
3D function with pixel intensity (Diffusion coeffecient of the gel) identify this is a bead
If the mask is not centered in the second largest well disqualified If the mask is too large, it’s likely just picking up the well.
In the future it would be interesting to mask
Failure modes
- No bead
- Failure in instrumentation
As a starting point:
- extracting mass statistics
- area of the mass
- pixel intensity
- What is the distribution of the summary stats
- Look at the far out ones and determine why they are like that
2025 02 28
- Graph helps pull out finer details from the image
- Eccentricity
- majour axis
- dot product between vector represetnation
- summary statics
- Deploy on hugging face
- Crop
- Focus on the more translucent beads
- Really fuocus!
- increase pixel intesnity
Real-time Classification
Parallel Processing and GPU Acceleration
Dynamically