基于多任务学习的自然图像分类研究

针对自然图像内容结构复杂、难以区分的实际情况,提出了一种基于多任务学习的自然图像分类方法。通过额外任务来辅助主任务的学习,构造了衡量任务间相关性大小的相关性矩阵,提出了主任务联合额外任务共同决策的学习模式;通过额外任务与主任务的相关性来控制额外任务参与主任务决策的程度,以提高主任务的分类准确率。实验结果表明,与传统的单任务学习相比,尤其是在已知样本较少的情况下,多任务学习机制能够明显地改善分类器的泛化性能。

Research of classification method fornatural images based on multitask learning

Considering it's difficult to distinguish different images because the content and structure of natural images is very complex, this paper proposed a classification method for natural images based on multitasks learning. It designed some extra tasks to assist the main task's learning, and constructed correlation matrix that measured the correlation between tastes. It proposed a study model that main task jointed extra tasks to make decision. And controlled the degree of extra tasks involved main task's decision though the correlation between main task and extra tasks. Experiment results show that the proposed method can markedly improve the generalization capability of classifier comparing with single task learning, especially in the case of lacking enough prior knowledge.