SUPA achieves 96% accuracy in film defect inspection for a global consumer electronics manufacturer

The Problem

• The Client, a global consumer electronics manufacturer, required SUPA to label images of film material at a manufacturing line, to identify and label defects including various types of tears and other features indicating catastrophic failure
• The project required domain expertise on the manufacturing process, and familiarity with film defects
• The Client required an accuracy of at least 90%

The Solution

• SUPA's Industrial subject matter experts worked with the Client to understand context and ML goals
• SUPA co-created the labeling instructions and provided feedback to clarify rules and edge cases – advising the Client to reduce the initial number of required labels from 5 to 3 while achieving the inspection objectives. This helped to reduce error rates while improving cost and time efficiency
• Before scaling the workflow, SUPA performed a POC with an initial small batch of data to verify that output matches the Client’s requirements for defect classification and polygon labeling
• SUPA deployed annotators with prior experience in labeling manufacturing defects to ensure accuracy is met

The Result

• Client data labeled successfully, with quality being consistently above 90%, and batch accuracies increasing up to 96% over time

‍