Wooden dowel defect detection

Our customer - the largest manufacturer and supplier of wooden dowels for the furniture industry, manufacturer and supplier of wooden dowels for the furniture, toy and brush industries.

Business problem

In a demanding wooden dowel production environment, their biggest challenge was dealing with a large number of visual defects on a wide range of their products.

• Existing production lines produced up to 1500 wooden dowels per minute.
• Up to 30% of produced wooden dowels had visual defects.
• Existing production lines required large mechanical or machinery changes to reduce defect rates.

‍

PROJECT OBJECTIVES

• Implement AI Computer vision solution without disrupting existing production lines or introducing big mechanical changes.
• Find a non disruptive solution to change faulty ejection system of the existing solution.
• Achieve false positive rates of <0.2% - overall number of defected dowels not recognized correctly.
• Achieve false negative rates of <5% - overall number of dowels without defects being recognized and ejected as defected dowels.
• Recognize defects as small as 2mm in size.
• Create a versatile solution, capable of recognizing visual defects in 40 different dowel types.
• Create a flexible solution with the ability to adjust accepted defect size for varying business needs.

‍

OUR COMPUTER VISION SOLUTION

Our advanced AI models were trained with thousands of images over 2 weeks time, to recognize possible defect types and achieve low percentage of false positives (defected items recognized as non-defective items).

High production speeds required out AI models to capture & analyze each image, and make a decision in as fast as 25 miliseconds.

Our skilled engineers spent days research and resolve existing design flaws in the machine's ejection system, to allow our AI models take control and achieve high levels of accuracy of the ejection.

RESULTS

We were happy to achieve all customer's requirements and improve the quality of their production line, reducing defect rates from around 30% to 0.2%.

We implemented ejection system changes, so ejection actions can be taken by the AI model.

We achieved stable and highly accurate defect detection and ejection results, reaching speeds, as fast as 1500 pieces/minute.

‍

‍