Semi-supervised Image Classification With Self-paced Cross-task Networks

Image classification,as is the basis of visual semantic understanding,has always been a concern in the field of computer vision.However,with the explosive growth of information and data,conventional machine learning algorithms,such as supervised learning and unsupervised learning,is not efficient to deal with big data problems.In detail,high cost of manual annotation is always a key issue in supervised learning and unsupervised learning is not effective to handle it.Therefore,many researchers pay their attention to semi-supervised learning,a learning paradigm that use labeled data and unlabeled data to train together.The research on semi-supervised learning started in the seventies of last century.Although semi-supervised learning has made great progress in theory,algorithm and practice,how to use the unlabeled samples effectively is still a core problem of semi-supervised learning.Traditional semi-supervised learning algorithms have been implemented with a variety of algorithms based on different assumptions and methods.However,the traditional semi-supervised learning algorithm has become increasingly unable to meet the development needs of the era of big data due to its limited ability to express itself,large memory overhead and high complexity of optimization and reasoning.Due to this reason,a self-paced learning based cross-task deep network is proposed for image classification in this dissertation.In a semi-supervised setting,direct training of a deep discriminative model on partially labeled images often suffers from overfitting and poor performance,because only a small number of labeled images are available,and errors in label propagation are,in many cases,inevitable.In this paper,we introduce an auxiliary clustering task to explore the structure of the image data,and judiciously weigh unlabeled data to alleviate the influence of ambiguous data on model training.For this purpose,we propose a cross-task network composed of two streams to jointly learn two tasks: classification and clustering.Based on the model predictions,a large number of pairwise constraints can be generated from unlabeled images,and are fed to the clustering stream.Since pairwise constraints encode weak supervision information,the clustering is tolerant of errors in labeling.Unlabeled images are weighted according to the distances to the clusters discovered,and a better discriminative model is trained on the classification stream associated with a weighted softmax loss.Furthermore,a self-paced learning paradigm is adopted to gradually train our deep model from easy examples to difficult ones.Experimental results on widely used image classification and pedestrian detection datasets confirm the effectiveness and superiority of the proposed approach.