Two-view Consistency-based Active Learning Algorithm In Image Classification

Data are the key to deep learning.Deep learning has made great success in image classification,where an important reason is the large amount of labeled data.Labeled open source image classification datasets,such as CIFAR-10,CIFAR-100 and ImageNet,have promoted the development of deep learning in academia.In actual industrial scenes,however,labeled data are still at a small number and most of the data are unlabeled.Data annotation is always expensive and time-consuming,and expert knowledge is even required in certain tasks like medical image analysis.Active learning and semi-supervised learning are two popular ways to reduce the burden of data annotation,of which the goal is to use less labeled data to obtain a higher performance.Active learning achieves this aim by actively selecting unlabeled instances to label,while semi-supervised learning by utilizing unlabeled data effectively,thus the integration of these two can yield more promising performance.In this article,we combine the active learning and FixMatch method together to form a more effective AL-SSL framework,and propose a novel consistency-based active learning query strategy called ?-Consistency,which calculates the inconsistency via the difference between no-augmented and augmented views of the same image.The idea behind our ?-Consistency strategy is that a good model should be robust,thus the output of the model should be similar even when the input instance is injected with noise,and the model is least robust for those samples have barely similar outputs,and learning these samples can improve the model's robustness.For each unlabeled instance,our method only needs to propagate forward twice,which greatly reduces the computational effort in the query process.We test the existing active learning approaches and our proposed approach on the CIFAR-10 and CIFAR-100 datasets.Experimental results show that our ?-Consistency outperforms existing consistency-based active learning methods on both CIFAR-10 and CIFAR-100 datasets.