Modality-Dependent Cross-Modal Retrieval Based on Graph Regularization

Table 2

The summarization of all compared methods.

Descriptions of comparison methods	Characteristics of comparison methods
CCA is a classic subspace method that projects different modalities into a common subspace to maximize the correlation between the paired information items.	Correlation analysis; Unsupervised learning
KCCA obtains the correlation between image and text through cross-view retrieval in a high-dimensional feature space.	Kernel correlation analysis; Unsupervised learning
SM projects image-text pairs into the semantic space to retrieve data from different modalities.	Semantic information
SCM projects an image-text pair to the semantic space in which learning is performed by CCA. SCM uses a combination of CCA and SM to improve retrieval performance.	Correlation analysis; Semantic information
GMLDA seeks the best projection direction so that the similar samples are as close as possible, and different classes of samples are as far as possible.	Generalized multiview analysis; Linear discriminant analysis; Semantic information
GMMFA combines semantic information, and CCA constraints to learn a common subspace through the combination of GMA and MFA.	Generalized multiview analysis; Canonical correlation analysis; Semantic information
MDCR performs different retrieval tasks for different query objects. Different projection matrices are learned to optimize each retrieval result.	Different retrieval tasks; Correlation analysis; Semantic information
JFSSL uses graph regularization to maintain similarity between intermodality and intramodality and performs feature selection for different feature spaces, thereby improving performance.	Graph regularization; Semantic information
JLSLR uses label graphs to learn the latent space and maintains a high correlation of multimodality features. The local relationships are maintained when different modal features are projected onto a common space.	Graph regularization; Semantic information
GSSSL learns a discriminative common subspace by combining the relevance of samples for different modalities with the semantic information.	Graph regularization; Semantic information
CRCMR not only uses dictionary learning to obtain collaborative representation for multimodal data but also takes into account the same semantic information of multimodal data.	Collaborative representation; Semantic information
CRLDA improves retrieval performance by considering the pairwise correlation between image features and text features and improving the discriminative characteristic of textual modality.	Different retrieval tasks; Correlation analysis; Semantic information; Linear discriminant analysis
CMOLRS adapts the margin of hinge loss for each triple, effectively utilizes sample features and semantic information and thus achieves a low-rank bilinear similarity measurement on data.	Relative similarities; Semantic information