A comparison of transformer and CNN-based object detection models for surface defects on Li-Ion Battery Electrodes

Deep learning-based defect detection approaches offer great potential for end-to-end surface defect detection. After the prevalent Convolutional Neural Network (CNN) models were state-of-the-art for almost a decade, transformer-based models recently surpassed the performance of CNN models on standard benchmark datasets. However, standard benchmarks such as the Common Objects in Context (COCO) dataset are not comparable to industrial use cases. To evaluate the applicability of transformer models in an industrial context, this paper applies a transformer-based object detection model for surface defect detection on Lithium-Ion Battery Electrodes LIBE and compares the results to a CNN-based object detection model. As a result, the transformer-based model outperforms the CNN model but is inferior in detection speed. In addition, the paper demonstrates the importance of a well-annotated dataset and shows the sensitivity of annotations for the model performance. Finally, this paper presents practical steps for an industrial application regarding backbone choice, inference speed, and metrics. • Deep-learning surface defect detection for lithium-ion batteries. • CNNs and transformers are both applicable to industrial object detection. • Computer vision as a key enabler for automatic optical inspection.