Semi-synthetic Data Set Based Lithium-Ion Battery Pole Piece Defect Inspection

A new method of classifying defective battery pole piece images was proposed in this paper. Firstly, a number of frequently repeating defects were identified in a data set provided by the industry, and subsequently manually labelled and classified. Then, a semi-synthetic data generation scheme was devised for each defect, thus solving the problem of sample imbalance and time-consuming data annotation. Next, a model with UNet architecture was selected to perform defect semantic segmentation reaching $7 7 \%$ mean dice score, and $6 6 \%$ mean IOU score on a manually labelled real validation set. Then, an inspection based on blob analysis is performed on the output of segmentation model to extract features for binary classification. Finally, an end-to-end tree boosting system XGBoost is trained and optimized to obtain sample-level binary classification. The experimental results show that the proposed method reaches precision of $94.4 \%$ and recall of $9 4.1 \%$ on the test set, which is an effective rate and can be a feasible solution to the Pole Piece inspection in a real-world scenario.