Classification of Metal Surface Defects Using Convolutional Neural Networks (CNN)

Abstrak

Metal surface quality inspection is an important step in ensuring that products meet predetermined industry standards. The manual methods used were often slow and prone to errors, so more efficient solutions were needed. The application of Machine Learning (ML)-based technologies, especially Convolutional Neural Networks (CNN), offered an innovative approach to overcome these challenges. CNN had the ability to automatically extract visual features from images with high accuracy, making it an effective tool in defect classification. This research used several CNN architectures, including MobileNetV2 and InceptionV3, as well as a model developed in-house, the K3 Model. Data augmentation, such as rotation and lighting adjustments, was applied to increase variation in the dataset and aid the model in generalization. The research results showed that the K3+Augmentation model achieved the highest accuracy of 100% in testing, with a very low loss of 0.0009. While MobileNetV2 offered better training speed, K3+Augmentation showed superior performance in detecting and classifying metal defects. These findings confirmed the potential of CNN in improving the efficiency of quality inspection in modern industry.