Multi-task Based Large-scale Image Classification And Application

In recent years,the image data on the Internet has grown rapidly,which triggered a huge demand for some application tasks like image annotation and image retrieval,and also brought new challenges for image classification,the base of these application tasks.For example,there are some serious problems like wrongly labeled,label missed,or data imbalance in the large-scale annotation image data provided from a variety of sources,which is not common in those previous small-scale classification tasks.To solve these problems,this paper has proposed the following work on the basis of the review of previous related work.A multi-task based large-scale image classification algorithm is proposed.The algorithm uses deep neural network as the basic framework,multiple classification tasks share the underlying neural network used for feature learning,and at the top,respectively,using their own independent classification network.This is based on all kinds of image sample error from different tasks in the top adjusting the weights of the whole neural network during the backpropagation process,and the purpose is to use the categories with massive data to help the feature learning of the categories with less data,but not to cover it,so as to improve the overall classification performance of all tasks.Especially,the paper analyzed the influence of the error weight from different scale tasks,and explored different weighting strategies.Besides,through the theoretical analysis,it is proved that the variation of the error weight is equivalent to the adjustment of the learning rate of the neural network,and the influence of the error weight on the network convergence performance is discussed through experiments.Experiments on large scale real data sets show the effectiveness of the proposed classification algorithm.In this paper,the classification algorithm is applied to an image retrieval system and a classification task for other data sets,which further verifies the effectiveness of the proposed algorithm.