Research On Image Classification And Video Tracking With Weakly Labeled Data

In today's society,artificial intelligence technology represented by deep learning is increasingly influencing people's lives.For many tasks,the successful application of deep learning relies on having large amounts of training data,labeled to a high standard.However,in many tasks,manual labeling of data is much more difficult than automatic data collection,which limits the application of deep learning technology to some extent.Compared to accurately labeling the data,weak labeling can greatly improve the efficiency of the work.Humans will inevitably receive some incompletely correct information in the process of continuous learning and progress,but humans can minimize the impact of such information on their own cognitive abilities.This paper believes that the deep neural network is robust,and its performance will not be rapidly degraded due to the presence of some weakly labeled data.At the same time,it can achieve performance similar to that of accurately labeled data by providing enough weakly labeled data.This paper first researches the impact of weakly labeled data on image classification.In the problem of image classification in this paper,weakly labeled data refers to the wrong labeling of the image category,which is the noisy label.The noisy label contains symmetric noise and asymmetric noise.For this kind of problem,this paper proposes a combination of ensemble learning and deep learning.A disagreement-based annotation method and different voting strategies are the core ideas of the proposed method.Through research in two noise environments,this paper proves the superiority of the proposed algorithm.This paper further researches the problem of tracking with weakly labeled data.Compared to the image classification problem,the labeling required by the tracking problem is much more complicated.Based on the research on image classification problems,this paper proposes a conjecture:on tracking problems,we can also use weakly labeled data to train neural networks.Therefore,this paper uses the weak tracker to provide a large amount of weakly labeled data as a training set to achieve the training of tracking neural network and applies this idea to the actual video surveillance scene.Experiments on public data sets prove the accuracy of the conjecture.Finally,this paper uses a combination of weakly labeled data and manual labeling on the self-built data set to further improve the performance of the neural network and proves the superiority and achievability of the proposed method.