An EfficientNet-Based Transfer Learning System for Defect Classification in Manufacturing

In semiconductor manufacturing industry, automated systems are essential for efficient and accurate identification of defects, prior to final product completion, to ensure quality and reduce waste. To achieve this, semiconductor industries are developing smart inspection systems to identify defects on the surface of wafers during manufacturing. Computer vision techniques play a crucial role in developing accurate inspection systems. However, most existing computer vision-based systems perform poorly when classifying defects, and many manufacturing companies still rely on manual inspection. To overcome this, we propose an efficient method for classifying defects in an industrial dataset using EfficientNet-B4 transfer learning along with Squeeze and Excitation block and multilayer perceptron. Furthermore, we applied data-augmentation techniques to enhance the dataset and improve the generalisation of proposed model. This proposed method is lightweight and can classify defects in real-time with an accuracy of approximately 98%.