Research And Application Of Deep Learning And Weak Supervision For Multi-Label Image Classification

With the development of network communication,collecting big dataset becomes possible.However,how to deal with these data efficiently is the most important issue.The increasing amount of data and powerful computational capabilities make deep learning a mainstream method in machine learning.And most successful methods are based on strong supervised information,which means that large amounts of data need to be annotated manually.Weakly supervised multi-label problem,which is an image contains a variety of objects without any of location information,is more general and closer to real life scene and has more practical value.These images are from unconstrained scene,due to the variety of background,light and deformation,the complexity of problem grow exponentially with multi labels.The traditional mainstream methods learn image representation through multi instance algorithm with hand craft feature.With the popularity of deep learning,there are some related works focusing on this problem.In this paper,we focus on the research and application of deep learning in weakly supervised multi-label image classification.The main contributions are as follows.First,the end-to-end structure of combination of deep learning and common multi label methods is proposed.Traditional multi label methods are divided into two steps: feature extraction and classifier.In the paper,we combine the deep learning with multi-label losses to form derivative structure and verify its effectiveness on public datasets.Secondly,we propose a weakly supervised multi-label image based on attention mechanism.In this framework,we combine convolutional neural network and recurrent neural network,make use of LSTM sequence learning ability,and add attention mechanism so that the model can focus on the part of the picture for the first time.The whole framework is end-to-end and trained with stochastic gradient descent learning,and automatically adjust the area of interest.Compared with the traditional depth-based learning method,performance is improved through our mechanism,and more importantly,it has stronger interpretability than other methods.Finally,we propose an imbalanced face recognition network architecture based on multi-task on specific task of face attributes recognition.Because of the problems of the imbalanced classes and the relation of labels,we use second order strategy and divide attributes into groups by relation and use a balanced strategy on face attribute recognition task.To a certain extent,the problem of imbalance is alleviated,and the network parameters are less and computational efficiency is higher.The performance of our method is state-of-art on the published large-scale face attribute data CelebA and LFWA.