Mastering the Mask: Harnessing Ensemble Learning for Defect Detection

The outbreak of the COVID-19 pandemic in 2019 led to a significant increase in the number of people wearing masks. As the demand for masks soared, numerous manufacturers entered the mask production industry. However, traditional quality control and defect detection processes in mask manufacturing heavily relied on manual inspection, resulting in substantial manpower requirements and time-consuming procedures. Although the implementation of AI-powered visual inspection aided in categorizing defects, distinguishing certain intricate flaws remained challenging, with an average model accuracy of 89.5% and an error rate of 10.5%. To address this limitation, this study adopts an ensemble learning approach, utilizing multiple models for prediction and employing majority voting to determine the final outcome. The ensemble learning technique successfully reduces the model's error rate by 16.2%, indicating its effectiveness in mitigating errors and enhancing accuracy. Nevertheless, solely relying on ensemble learning does not sufficiently lower the error rate. Thus, this study further investigates model ensemble by adjusting model weights to assess whether accuracy can be further improved. Experimental results demonstrate that with weight adjustment, the error rate can be minimized to as low as 7%, representing a two-thirds decrease from the current average error rate. This substantial improvement signifies the potential of weight adjustment in enhancing mask defect detection accuracy.