Vision Algorithm
This section covers the rules and models that pick defects out of an image, and the cost when a decision is wrong: how rule-based and deep-learning methods share the work, anomaly detection, and filter order. It is the stage that turns the image made by lighting and optics into a pass/fail decision.
Start with these
- Deep Learning vs. Rule-Based Algorithms, When to Use Which — the conditions that favor each
- Overdetection or Missed Detection, Which One Costs More? — setting the decision line by cost
- Building a Vision Inspection App with PySide6 and OpenCV — full code for a minimal UI that tunes parameters on screen
- Must Every Normal Patch Be Remembered? — coreset design for memory-bank anomaly detection
Unfamiliar terms are explained in the Glossary, and a reading order across topics is in the Start Here guide.
-
Does Memory-Bank Anomaly Detection Need to Remember Every Normal Patch?
Storing every normal patch takes 6.29 GB and a 1 % coreset 62.9 MB, yet search still costs 2.58 trillion multiply-accumulates per image. Search time caps the coreset ratio, and overdetection on normal…
-
Why Does a Lightweight Model Lose the Finest Defects First?
A 30 μm trace nick, 3 px at 10 μm/px, drops to 0.75 of a cell at student output stride 4 and can vanish whatever the distillation temperature. Stride 2, T = 2…
-
In Edge Inspection, Input Drift Shows Up Before Accuracy Drops
Lens contamination and light degradation show up in reference luminance and sharpness before accuracy drops. A per-frame p99 alarm fires once every 100 frames, so thresholds go on λ = 0.05 EWMA values…
-
One Lighting Angle Decides the OCR Read Rate on Molded Characters
Molded characters 50 μm high with 300 μm stroke spacing read by their shadows only between 18.4° and 53.7° of elevation. Choosing 20° keeps them readable down to 13.4 μm of height, but…
-
Why an Average Latency Inside the Deadline Is Not Safe
An average of 33.6 ms is just over half of a 60 ms deadline, but per-stage p99 values add up to 53.1 ms, leaving 6.9 ms. Transfer and inference each add 6 ms…
-
Where Do Fine Defects Go When You Shrink to 640 × 640?
Shrink a 20 MP PCB image to 640 px and a 90 μm defect is left at 0.35 px. Seventy tiles with a 64 px overlap take 560 ms, past a 400 ms…
-
Adding Reasons to an AOI Defect Call with a Vision Language Model
A vision language model behind AOI can only cite differences the lighting has already captured. Two channels of coaxial and low-angle ring light, ten pixels on an 80 μm bridge, 130 human reviews…
-
Catching 20 μm Sink Marks on High-Gloss Molded Parts with Phase Measuring Deflectometry
A sink mark 20 μm deep and 3 mm wide tilts the surface by 20.9 mrad, which moves the reflected fringes 16.8 mm on a screen 400 mm away. A 2.9 mm defocus…
-
The Error Frame Temperature Leaves in Two-Sided Chromatic Confocal Thickness Measurement
A 1 K change in the steel frame between two heads moves the thickness reading by 1.17 μm. How two-channel, two-sided chromatic confocal measurement removes the refractive index, and how heat, tilt and…
-
Camera Interface Bandwidth Determines Resolution
Running a 12 MP 8-bit camera at 60 fps takes 737 MB/s, not 2.2 GB/s. From 1GigE to CoaXPress 2.0, this works out the pixel size each link allows, what happens to a…
