Electrode Segmentation in Battery Cells with Gradual Learning

This work investigates one-shot segmentation of anodes and cathodes in computed tomography volumes of a battery stack cell and a cylindrical cell. To achieve this, Gradual Learning is applied, which relies on the high similarity of neighboring slices. After training, pseudo labels of these neighboring slices are generated and repeatedly used for retraining. A critical hyperparameter in this process is the step size n, which refers to the number of neighboring slices added in each iteration. The results demonstrate a clear trade-off between achieved segmentation performance and execution time, influenced by this parameter. For the stack cell, the Intersection over Union increased from 0.731 (baseline trained without Gradual Learning) to 0.819 for a small n = 25 and to 0.769 for a large n = 75. The performance gain was smaller for the cylindrical cell, increasing from 0.779 to 0.784 for a large n = 15 and to 0.789 for a small n = 5.