Visual Inspection System for Manufacturing
A visual inspection system that compares a reference image against an inspection image to detect scratches and defects. Preprocessing, alignment, and a multi-model fallback reduce missed defects.
Duration
Contract development
Team
1-2 people
Client
A manufacturer
01 // OVERVIEW
Project Overview
A visual inspection system for manufacturing that detects scratches and defects from product images. Preprocessing and alignment absorb variation in shooting conditions, and an AI model outputs a list of defects with bounding boxes. Falls back to a different model depending on confidence level.
02 // KEY FEATURES
Key Features
Preprocessing and Alignment
Resizing and CLAHE color correction normalize shooting conditions, and SIFT aligns the image with the reference. This makes comparisons robust to lighting changes and subtle color differences.
Multi-Model Fallback
A fast model makes the initial judgment, and only when confidence falls below a threshold does a different model re-evaluate, balancing speed and accuracy.
Defect List and Location Output
Outputs the type and bounding box of each detected defect, visualizing the inspection results.
03 // TECHNICAL DEEP-DIVE
Technical Highlights
Solution
Countermeasures Against Missing Fine Scratches
Higher resolution, use of difference detection, and prompt optimization per defect type continuously improved the detection rate for small scratches and resilience to noise.
04 // IMPACT & RESULTS
Results
Multi-Stage
Confidence-based fallback
Location Output
Defect sites visualized