Computer Vision
Computer Vision
What Is Computer Vision?
- Computers 'seeing' and understanding images and video
- The eyes behind face recognition, defect detection, and self-driving
- CNNs are the flagship technology
What computer vision really changes is 'consistency of judgment.' Human inspectors apply different standards in the morning and the afternoon, and their calls vary with experience. Vision inspection gives the same answer for the same image and keeps every photo and coordinate behind each verdict. So the first benefit is not 'fewer people' but being able to trace where and why defects slipped through. Inspection records that build up as data become the evidence for equipment improvements and for responding to client claims.
In practice you will meet roughly five forms: classification, which labels a whole image; detection, which also draws a box around the location; segmentation, which outlines regions pixel by pixel; OCR, which reads text; and anomaly detection, which learns only normal patterns and flags anything that deviates. On lines with only a few dozen defect samples, anomaly detection is more realistic than the first three. The requirements differ too. A detection model needs hundreds to thousands of labeled images per defect type, and its performance drops sharply when lighting, angle, or background changes. Where processing happens is also a choice: if the line moves several units per second, use an edge setup that finishes within tens of milliseconds on on-site equipment instead of a round trip to the cloud. A multimodal AI that describes images and a dedicated vision model that judges a 0.2mm scratch are tools for different jobs.
The most common mistake on site is the belief that you just need to install a camera. Half the project effort goes into lighting and jigs, that is, making sure every shot is taken under the same conditions. The second is the accuracy number. On a line with a 0.5% defect rate, a system that calls everything good still scores 99.5% accuracy. The numbers to watch are the miss rate (the share of defects missed) and the false-reject rate (the share of good items bounced). Each miss turns into a client claim, and a spike in false rejects means scrapped good products and inspector rework. The third is the illusion that you train once and you’re done. When part specs or material suppliers change, detection rates can collapse within months, so unless quarterly retraining and performance checks are written into operating procedures, the expensive camera ends up switched off and abandoned.
John KimInstructor John Kim, author of "B2B Marketing Design" and CEO of Nanum Consulting
Sungil KimGlobal pitching expert from KAIST and Samsung Electronics who coached the winning teams at the UN Tourism and Select USA IR pitch competitionsA field of research since the 1960s. In 2012, deep learning (AlexNet) sharply improved image recognition performance and accelerated real-world adoption.
Consider a hypothetical case. Company A makes auto parts: a tier-2 supplier with 180 employees and KRW 42 billion (≈ US$31,423,000) in annual revenue. Twelve inspectors across three shifts checked appearance by eye, and last year 9 missed defects reached customers, with line-stoppage compensation on top. The quality assurance team took on a six-month project to introduce vision inspection.
- Manager Park on the quality assurance team gathered 1,200 defect photos from three months and found the labels were all over the place. After hearing that 'inspectors had logged scratches and dents differently,' the team rebuilt the reference photo album for judging defects first.
- The first trial model reached a detection rate of only 71%. The diagnosis was that lighting, not the algorithm, was the cause. After the equipment engineering team added ring lights and fixed jigs to standardize shooting conditions, the rate rose to 89%.
- With only 40 defect samples per type, the team switched to an anomaly detection approach trained on just 8,000 normal parts. Acceptance thresholds of a miss rate of 0.3% or less and a false-reject rate of 5% or less were written into the vendor contract.
- For two months, three inspectors double-checked every AI verdict. Once the line manager set the rule that 'people only recheck items flagged as false rejects,' daily rechecks dropped from 220 units to 60.
- The team made retraining every quarter on 300 new defect images a standing item on the quality meeting agenda. In Q4, when the material supplier changed, they caught a 6-percentage-point drop in detection rate within two weeks.
- A year later, defects missed and leaked to clients fell from 9 to 1, and inspection time per unit dropped from 22 seconds to 4. Six of the 12 inspectors were reassigned to assembly and packaging.