Purpose
Recover a more stable view from a short sequence of imperfect captures.
Role in the complete pipeline
Stacking sits between enhancement and OCR when a single frame is not dependable enough.
INPUT
→A short aligned burst of page crops.
METHOD
- Collect a bounded frame burst
- Estimate and correct small alignment shifts
- Take a median value across corresponding pixels
- Compare the result with the best single frame
A median-stacked candidate and the strongest single-frame alternative.
Common failure cases
- Frames are too misaligned
- The page moves during collection
- Glare remains fixed across every frame
- Excessive depth adds delay
How VisoraAI responds
- Reduce the burst depth
- Use the best single frame
- Request a steadier capture
Current limitations
- Stacking is not guaranteed to improve recognition
- Alignment and processing time constrain depth on edge hardware
Planned improvements
- Measure the depth/latency trade-off
- Benchmark against Character Error Rate
- Tune alignment for Raspberry Pi