Semi-supervised Active Learning Based on Semantic-aware Crop Consistency

Reducing labeling expenses is the main goal of active learning. However, due to the problem of cold start, the poor sample selection of the initial pool decreases the performance of many active learning methods. Besides, pure active learning methods fails to utilize the information of the unlabeled samples for model training. To tackle the problems, we combine active learning with consistency regularization based on cropping of the semi-supervised learning. To reduce the possibility that random crop may lead to false positives and thus poor understanding of the cropped content by the model, we use a localization module that is semantically aware to significantly lower false positives. The combination of both semantic-aware crop consistency-based semi-supervised learning and active learning can maximize the classification performance with minimal human cost. The method is evaluated using the three distinct image classification datasets CIFAR-10, CIFAR-100, and SVHN. The results illustrate the superiority of our method over competing methods.