Industrial Product Defect Detection Using Custom U-Net

Numerous automated labour-saving systems have been created and implemented to lower production costs and enhance product quality. Systems for intelligent visual analysis are now playing a more significant role in production lines. Many deep learning and machine learning techniques were previously used to identify defective products. Still, the models were never tested because they did not produce satisfactory results and had numerous other problems adjusting to sparse and poor-quality data. To address this issue, a deep learning-based customised UNet model was introduced in this proposed work. This model was trained on six different classes of data-sets to assess the model’s capability on different image textures and resolutions. The proposed model had an overall accuracy of 97.25%. We also achieved significantly higher precision, recall, and F1 scores of 96.04%, 95.58%, and 96.12%, respectively. The execution times of six different classes can indicate how well a model performs, and the average execution time is only 33.1 seconds, according to observations.