Application Research Of Multi-criteria Active Sampling Strategies On CNN-based Image Annotation

With the success of the deep learning applied in the image recognition field,there has been a great leap in the performance of image classification.Different from the traditional neural network,deep learning algorithm simulates human brains to build a deep network structure as well as employ a variety of training skills,so it could leam complex functions to generate robust image features.In generally,the performance of image classification is proportional to the number of layers in networks,and training a robust network model requires a larger number of labeled samples.In the digital age,the acquisition of massive images is convenient,but labeling a large number of images is laborious and sometime may introduce noise.In order to save labeling costs more effectively,active learning is proposed.Active learning aims to select a small amount of the most informative samples from the unla-beled set for manual labeling,so that the learned model based on the labeled samples could achieve a good performance.In active learning,the sample selection strategy is a key re-search issue.The common sample selection strategy only considers one criterion or two criteria like uncertainty,diversity and minimum error etc..This onefold strategy plays a limited role in deep learning algorithm,and its utility will be gradually exhausted as the training goes on.For the problem that the onefold active learning strategy is not suitable for deep learn-ing,this paper studies how to integrate multiple criteria to select the most informative sam-ples to speed up the performance improvement of deep learning algorithms.The main research work and innovations of this paper are as follows:firstly,we propose a novel multi-criteria active sampling strategy for deep learning called LEVER(short for "muLti-critEria actiVe dEep leaRning")that considers density,similarity,uncertainty,and label-based measure simultaneously.The density and similarity are used to reduce the duplicated information between labeled samples;The uncertainty helps us accelerate the convergence of the model;and the label-based measure could speed up model performance improvement as well as make a performance balance among classes.Besides,to maximize the advantages of multiple criteria,this paper designs an adaptive fusion methodour method,which adap-tively adjusts the weights by exploring the utility of each criterion at different training stages.Secondly,this paper designs a novel active learning strategy to avoid the performance im-balance among classes in deep learning,which is called "the label-based measure".In the early stage of training,the strategy focuses on the class with the fastest performance to improve the performance of the model.In the late training period,the strategy focuses on the poor performance of the class to balance the development of the model.Therefore,it can effectively alleviate the performance imbalance among classes.Thirdly,we conduct a series of experiments on the MNIST and CIFAR-10 datasets,and the experimental results demonstrate the superiority of LEVER.