【摘要】 空间信息表示是增强图像特征表达性能的重要手段,通过空间关系建模与深度学习方法融合可有效提升深度特征的语义特性,从而提升图像检索性能。首先,针对复杂图像的空间关系表示提出了一种新的精细拓扑结构表示模型,该模型不仅具有完备的拓扑描述性能,还提供了两种拓扑不变量的推理算法,使得拓扑不变量可以由表示模型直接推导而不需要繁复的几何计算;其次,基于精细拓扑结构表示模型,提出了有效的拓扑结构相似性度量方法,为空间关系特征表达奠定了基础;最后,进一步结合卷积神经网络,提出融合复杂空间关系特征与深度特征的多目标图像检索方法。实验结果表明,所提出的拓扑结构表示模型在空间查询中具有良好的性能;所提出的图像检索框架取得优于现有方法的精度,并能够有效地结合手工特征与深度特征各自的优势,为提升深度学习方法的可解释性创造了有利条件。更多还原

【Abstract】 Spatial information representation is an important means to improve image visual feature representation performance. The integration of the spatial relation model with deep learning can effectively enhance semantic property of deep features, improving the image retrieval accuracy. In this paper, we proposed a novel detailed topological structure representation model to describe spatial relation of complex images. This model not only had complete topology description performance, but also provided two efficient reasoning algorithms, which made the topological invariants directly deduced from the model without any geometric calculations. Similarity matching approaches based on fine topological structure representation model was proposed for spatial relationship feature representation. Finally, in combination with convolution neural network, a multi-object image retrieval framework was developed by fusing the spatial relation features and deep features. Experimental results demonstrate that the proposed topological model has remarkable performance in spatial query. Moreover, the proposed image retrieval framework outperforms the current methods in terms of precision and with advantages of both the manual and deep features, it provides a superior means to improve interpretability of deep learning methods.更多还原