| Nowadays,the visual reasoning system has been able to answer complex questions that cannot be answered by the visual question answering system on the CLEVR dataset.However,the current complex visual question answering system has some defects,including the problems of model over-fitting,high labeling cost,and poor generalization ability caused by strongly supervised learning;the system is difficult to deal with the long problems caused by the structural defects of the model.;the problem of the system getting into the local optimal solution caused by the greedy algorithm.To solve the above problems in the complex visual question answering system,this paper first analyzes and discusses the reasons for the above problems.Then,the improved methods are proposed to solve these problems: 1)the attention mechanism was introduced in the program generator,which improves the ability of the model to deal with long problems and improves the accuracy of the model;2)the beam search algorithm was introduced in the program generator to enable the model to jump out of the local optimal solution and find a better solution,which improved the model performance on the CLEVR data set;3)proposed a complex visual question answering system based on active learning,which makes use of the characteristics that active learning can select training samples efficiently and accurately,and selects the most informative and representative samples as training sample set,so as to achieve the goal of training a more accurate complex visual question answering system with fewer samples.Finally proved by experiments: 1)Improved effectiveness based on attention mechanism;2)Improved effectiveness based on beam search algorithm;3)The effectiveness of the complex visual question answering system based on active learning,and the comparison and verification results from three aspects show that combining active learning is very effective for the complex visual question answering system. |
PDF Full Text Request