Visual Quality Control

Detecting a possible defect and deciding what to do with the unit are two separate steps. The confidence threshold determines whether the product ships, goes back for rework, or needs a person to inspect it, while the cost of a missed defect helps the business decide where that threshold belongs.

Each sample unit is generated with either a known defect or no defect, which makes it possible to compare the system's decision with the correct outcome. The system compares the color and brightness of every bulb with the expected result, then combines the overall difference with the single largest defect to calculate one confidence score. A clearly incorrect bulb can therefore lower the result even when the rest of the string looks correct.

A separate rule treats three or more consecutive dark bulbs as a wiring fault and sends the unit to a person regardless of confidence.

This project connects a replica of AI-enabled visual inspection with an automated routing decision. The confidence score represents what a computer-vision system could detect, while the resulting action shows how an automated workflow could send the product to shipping, rework, or manual inspection.

The confidence score and wiring-fault rule remain separate so the user can see which signal caused the decision. A detected wiring fault always goes to a person regardless of the threshold.

The demonstration includes 16 sample units and five defect types. A production system would train on labeled images from the actual manufacturing line and would also account for changes in camera angle, lighting, and product position.

A production version would connect a trained computer-vision model to labeled images from the actual line, then route each unit to shipping, rework, or manual inspection based on the result.

This project uses TypeScript in the browser to compare each light against the expected result and calculate a confidence score based on the differences it finds. At a larger scale, an AI-enabled camera and trained computer-vision model could use the same process to inspect each unit and determine what should happen next.