Research On Deep Learning Based Image Recognition With Noisy Labels

Benefiting from the dramatic increase in computing power and the rapid growth of training data in recent years,deep learning has become the mainstream algorithm for many computer vision tasks.However,collecting a large scale dataset with clean la-bels is expensive and time-consuming.On one hand,expert knowledge is necessary for some datasets.On the other hand,we can easily collect a large scale dataset with noisy annotations from various websites.But such noise makes networks overfit seriously and accuracies drop dramatically.Therefore,finding an efficient and robust algorithm to handle datasets with noisy labels has a great research significance and practical ap-plication demand.This paper explores the problem about noisy labels based on deep learning,the main contributions are summarized as follow:1.Research on single-label image recognition with noisy labels.In the noisy la-bel problem,single-label image recognition is the most important issue and has long been researched.At present,many methods require additional auxiliary in-formation,including additional clean dataset and ground-truth noise transition matrix.In addition,many methods based on designing robust loss do not per-form well on real-world datasets.Therefore,we propose an end-to-end frame-work called PENCIL(probabilistic end-to-end noise correction in labels),which can update both network parameters and label estimations as label distributions.PENCIL is independent of the backbone network structure and does not need an auxiliary clean dataset or prior information about noise,thus it is more general and robust than existing methods and is easy to apply.In experiments,PEN-CIL outperforms previous state-of-the-art methods by large margins on various datasets.2.Working principle and extended application of PENCIL.PENCIL adopts la-bel probability distributions to supervise network learning and to update these distributions through back-propagation end-to-end in every epoch.Thus,the up-date process is very important,and the selection of the classification loss function in the PENCIL framework will have a great impact.In this paper,we propose an inverse KL-divergence,which is different from previous methods but is robust for noisy label handling.Then we show that the inverse KL-divergence is indeed more suitable than the original KL-loss and the mean square error in our pro-posed PENCIL framework.Finally,based on the working principle of PENCIL,we find that repetitive training of PENCIL can achieve better performance.3.Research on multi-label image recognition with noisy labels.When we ex-tend the problem from single-label to multi-label,the complexity of the problem increases a lot.Therefore,the original network structure together with sigmoid function is too simple to handle this situation.Our PENCIL framework corrects noisy labels based on the predictions provided by the backbone network.There-fore,a backbone network with better performance of multi-label classification tasks is very important.Thus,we propose a simple attention structure to replace the global average pooling layer in the original backbone network.The attention structure can be applied on any network,which retains the advantage of PEN-CIL that is independent of the backbone network structure.In the experiments,performance of networks with different sizes with the attention structure obtains significant improvement.Our proposed PENCIL framework is also effective on multi-label tasks after combining with the attention structure.