一种应用多储备池回声状态网络的图像语义映射研究

Research on Image Semantic Mapping with Multiple-Reservoirs Echo State Network

【目的】建立图像低层特征到高层语义的映射,填补图像检索中的"语义鸿沟",以提高检索准确率。【方法】借鉴集成学习思想,将多储备池回声状态网络(MESN)应用于图像语义映射模型中。图像低层特征按照类型划分后,通过不同的储备池训练,并对训练结果进行线性融合。【结果】该模型相对于BP神经网络和传统ESN,平均映射错误率分别下降31.64%和19.28%,查准率分别提高4.56%和1.86%。【局限】储备池参数通过人工设定,未构造参数优化算法。【结论】实验结果证明,将多储备池回声状态网络应用于图像语义映射中是有效的。

[Objective] The mapping between low-level visual feature and high-level semantic information is built up to fill the ＂semantic gap＂ of image retrieval and improve accuracy. [Methods] Referring to the idea of ensemble learning, Multiple-Reservoirs Echo State Networks （MESN） is applied to semantic mapping model. After the low-level visual features of images are divided by feature types and trained by different reservoirs, the training results are combined linearly. [Results] Compared to BP Neural Network and traditional Echo State Network, the average error rate of MESN decreases by 31.64% and 19.28% respectively, the precision rate increases 4.56% and 1.86% respectively. [Limitations] The parameters of reservoirs are set artificially. Parameter optimization algorithm isn＇t constructed. [Conclusions] Experimental results show that the semantic mapping model of Echo State Networks with Multiple-Reservoirs is effective.