Figure 1.

Workflow of the proposed approach. A deep CNN model is first trained with labeled data samples. The trained model is then evaluated on unlabeled data to generate pseudolabels for the unlabeled data. A pseudolabel selection algorithm that integrates a class balancing mechanism is used to select pseudosamples that have the highest confidence probability confidence score. The selected samples together with their pseudolabels are used to augment the training sample for the next training iteration and the cycle is repeated iteratively until a stopping criterion is met.