Research On Deep Active Learning Technology Based On Unsupervised Feature Matching

Existing works show that the supervised training effect of deep learning models is directly proportional to the amount of labeled data,but the cost of labeling massive data is often very expensive,so the limited labeling dataset has become one of the main bottlenecks in the current deep learning performance improvement one.Although Weakly Supervised Learning,Unsupervised Learning,and Contrastive Learning are expected to solve the problem of expensive data labeling,the current performance of such technologies is still lower than that of supervised training.Active Learning aims to sample the most valuable data from massive unlabeled data for labeling and training the network model.This technology can maximize the training effect of the model under the limited labeling cost and thus can avoid expensive labeling costs.In the context of big data and deep learning,active learning technology has increasingly become a hot research direction.The main content of this paper is the research of deep active learning technology based on feature matching,and more specifically,the use of Feature Matching to alleviate the data bias problem in active learning.However,the existing feature matching and active learning have the following problems to be solved urgently: First,the existing feature matching technology extracts a small number of features to cause the lack of key information or extracts a large number of features to cause a large amount of irrelevant information redundancy.At the same time,feature matching based on supervision methods is not only costly but also poor in generalization ability.Second,the existing active learning techniques,especially those uncertainty-based active learning methods,are prone to data bias problems,that is,the obtained sampling results tend to be biased toward the partial classification of the data set.Third,the existing active learning techniques can only be oriented to simple visual tasks such as classification,and how to generalize the existing technology to more complex visual tasks is also a current problem.Faced with the above problems,this paper proposes a deep active learning framework based on unsupervised feature matching.The specific research content and contribution include:1.This paper proposes an unsupervised feature matching technology,which ensures that the information of the target to be matched is obtained comprehensively and completely by extracting the features of different network layers of the model and fusing;at the same time,considering that the information of some feature channels is redundant or invalid Yes,this article filters the channels purposefully based on the variance information to reduce the noise of the selected features.In short,the unsupervised feature matching technology proposed in this paper is characterized by the fusion of multi-layer features and the selection of feature channels with rich information.On the public data sets Oxford 5k and Paris 6k,the unsupervised feature matching method proposed in this article has a 1%-4%accuracy improvement over similar technologies in multiple feature dimensions(including 128-dimensional,256-dimensional,and 512-dimensional).2.The existing active learning technology has the problem of data bias.The fundamental reason is mainly that there is redundant data in the sampled data,that is,there are similar data.To this end,this paper proposes a new active learning framework that uses the proposed unsupervised feature matching technology to compare features of selected uncertain data and remove some of the data with similar features,thereby achieving the mitigation of data bias.Purpose.In order to verify the proposed active learning framework,this paper has been widely used in classification and detection tasks.The data sets used include CIFAR-10,Fashion-MNIST,PASCAL VOC 2007,PASCAL VOC 2012,etc.The baseline model models used include Inception V3,Res Net-50,Mobile Net V3,etc.,and the baseline detectors used include Efficient Det,Faster R-CNN,SSD,etc.The experimental results show that the active learning framework proposed in this paper is better than the existing active learning technology under the above experimental settings.At the same time,techniques such as feature visualization and data sampling point visualization also prove that the active learning framework proposed in this article can effectively alleviate the problem of data bias.3.Least Confidence,Edge Sampling,Entropy,and other methods are all existing typical active learning uncertainty sampling techniques,but the sampling basis of such techniques is based on category information,So it is often only suitable for classification tasks.For the target detection task,in addition to the category information,the position information of the object to be identified is equally important.Therefore,how to take the position information into consideration in the uncertainty sampling process has obvious gains for the target detection task.Based on the above considerations,this paper improves the existing uncertainty sampling technology for classification tasks,that is,considers both the category information of the object to be identified and the location information of the object.Experiments on target detection tasks show that the uncertainty sampling method in this paper is better than the existing technology.