Cross-modal image–text retrieval serves as a fundamental bridge between visual perception and linguistic cognition. However, the inherent semantic entanglement of visual data poses significant challenges: unstructured visual content typically encapsulates holistic semantic information, whereas structured textual descriptions exhibit inherent subjectivity and locality. Existing paradigms struggle to capture this asymmetric cross-modal information, ideally yielding only averaged representations that suppress fine-grained details. Although set-based embedding methods attempt to alleviate this asymmetry, existing approaches often lack guidance from explicit structured semantics and focus solely on local view optimization, leading to stochasticity in semantic subspace decomposition and gradient sparsity. This paper proposes Multi-view Semantic Topology Alignment with Reasoning from VLMs (M-STAR), a multi-view modeling framework. We design the Semantic Prior Acquisition via MLLM (SPAM) module, employing a dual-granularity semantic encoding mechanism to guide the structured reconstruction of visual features. Furthermore, we introduce the Multi-View Global Instance Discrimination (MGID) module, which imposes cross-view discriminative constraints via a centroid pool, ensuring the uniqueness and robustness of visual embeddings within the global semantic space. Finally, we propose the Holistic Multi-View Alignment (HMA) optimization strategy, facilitating collaborative learning across subspaces and significantly enhancing inter-view diversity. Extensive experiments on the Flickr30K and MS-COCO datasets demonstrate that M-STAR achieves state-of-the-art performance and exhibits superior robustness in handling complex semantic scenarios.

Artistic style transfer synthesizes a stylized image that preserves content geometry while substituting texture, color, and stroke statistics from a target artwork. Most neural style transfer models are designed and evaluated under the visual priors of Western paintings, where sharp edges, high-contrast textures, and explicit contours broadly define the style distribution. Directly applying them to traditional Chinese landscape painting often exposes a core conflict: maintaining globally coherent scene structure (e.g., mountain ridges, river trajectories, spatial depth) while producing region-aware abstraction with meaningful brushwork. This gap commonly leads to incomplete stylization, inconsistent perspective, or structural distortion, particularly in areas that require long-range continuity. To address this challenge, we present DynBrush, a quality-efficient style-transfer framework specialized for Chinese landscape rendering. DynBrush adopts sequential spatial modeling with a decoupled dual-channel encoder that disentangles structural and stylistic representations. We introduce Structural State Evolution (SSE) to reinforce global layout and geometric stability progressively. Stylistic State Refinement(SSR) complements this process by enhancing chromatic embedding and ink-tone transitions to match ink-and-wash aesthetics better. For decoding, Structure-Guided State(SGS) anchors reconstruction to globally consistent structural cues, while Cross-domain Style Attention(CSA) improves regional style correspondence across appearance discrepancies. We further incorporate Blank-Preserving Multi-scale Context Enhancer(BMCE) to organize brushstroke hierarchies across scales, yielding coherent stroke granularity, stable ink-tone gradation, and reserved-white fidelity. Experiments demonstrate that DynBrush delivers stable structure preservation and style consistency across diverse landscapes, with a favorable quality–efficiency trade-off and strong adaptability to traditional Chinese landscape characteristics.

Unsupervised cross-modal hashing retrieval has demonstrated significant advantages in heterogeneous retrieval tasks, as it requires no manually annotated data and exhibits strong scalability across modalities. However, existing unsupervised methods often suffer from an imbalance in the representation quality between image and text features. Moreover, the lack of explicit semantic supervision frequently leads to semantic degradation, which limits retrieval accuracy. To address these challenges, we propose a novel unsupervised cross-modal hashing retrieval framework, termed Contrastive Multimodal Fusion and Pseudo-Labeling (CMFPL). Specifically, we design an efficient contrastive multimodal learning module that integrates a Transformer-based multimodal encoder with contrastive loss. This design not only enhances deep semantic interaction between modalities but also improves the expressiveness and discriminative capability of modality-specific features. In addition, we develop an innovative multimodal association matrix generator, which combines cosine similarity with K-Means clustering to more accurately model the semantic correlations across modalities. This matrix is further utilized to guide a Graph Convolutional Network (GCN) for semantic reconstruction of hash codes. To reinforce semantic supervision during the hash learning process, we leverage the zero-shot classification capability of CLIP to generate pseudo-labels for each sample and construct a high-quality semantic association graph. This graph provides robust semantic guidance for hash code learning, thereby significantly improving retrieval performance. Extensive experiments conducted on three publicly available cross-modal datasets demonstrate that our proposed CMFPL framework consistently outperforms existing state-of-the-art methods in terms of retrieval accuracy and robustness, fully validating its effectiveness and superiority.

The main challenge in image-text matching lies in bridging the gap between visual and linguistic modalities for accurate cross-modal semantic alignment. While current mainstream methods enhance local feature interactions through region-word attention mechanisms, their isolated object modeling paradigm fails to capture deep semantic relationships, limiting fine-grained cross-modal reasoning capabilities. Research has explored structured relationship models like scene graphs; however, the visual modality faces inherent limitations: unlike text, which can build relationship graphs through lexical logic, visual scenes lack clear contextual semantics, resulting in issues such as ambiguous entity boundaries and distorted relationships that impede effective cross-modal alignment. This paper proposes a VLMs bridging-enhanced Scene Semantic Reasoning framework (VSSR). Based on modal characteristic differences, we construct a dual-path scene parsing framework: On the visual modeling, leveraging the strong semantic understanding capabilities of Vision-Language Models (VLMs) to generate dense scene semantic labels; On the textual modeling, fully exploiting the structured advantages of language by designing a graph attention network guided by relationship inductive bias to deeply mine implicit semantic associations between textual entities. To further bridge the modality gap, we create a multimodal collaborative representation space, using scene semantic labels as anchors to bridge the two modalities and achieve cross-modal knowledge transfer through joint semantic projection. While maintaining linear computational complexity, this architecture realizes fine-grained matching from scene-level (image-caption) to entity-level (image-entity) through relation-aware semantic modeling. Experiments on the Flickr30K and MS-COCO benchmark datasets demonstrate that VSSR outperforms existing state-of-the-art approaches in retrieval performance.

In recent years, remarkable progress has been made in remote sensing image-text retrieval (RSITR), which has transitioned from relying on compact global features to more fine-grained local features representing salient objects in images. However, existing methods typically concentrate only on significant entity information in remote sensing images as local features, overlooking the correlations between entities, resulting in isolated entity information. Moreover, there is ample room for exploring text local features.To address these issues, this paper presents an Entity Spatial Relation enhancement Network (ESRN), leveraging global and local entity features in remote sensing images and texts. For local feature processing, Graph Convolutional Network (GCN) is employed to aggregate the correlation between entity features and entity information in remote sensing images, enhancing entity information learning. The self-attention mechanism is used to model the remote dependency of entity keywords and spatial orientation semantic relations in text, strengthening the text representation ability. A strategy of proportionally adding different levels of features is proposed to enhance the representation of salient features and reduce noise interference.The approach was evaluated on two renowned remote sensing datasets, RSICD and RSITMD, validating the model's capacity to perceive the semantics of remote sensing images and text entities. Performance comparison, ablation experiments, and visualization analysis convincingly demonstrate the state-of-the-art performance of the ESRN method in the RSITR task.

Abstract Although the fact that current methods have some effects, unsupervised cross-modal hashing methods still face several common challenges. First of all, the text features that have been collected from text data are not comprehensive enough to provide sufficient guidance for building textual modal similarity matrices. Secondly, the fusion of similarity matrices from different modalities lacks adaptability, leading to a less accurate final similarity matrix. This work suggests Enhanced Similarity Attention Fusion Hashing (ESAFH) as a remedy for these problems. Firstly, we construct a text encoder to enrich text features, an adjacency matrix is built to represent the association relationship between pairs of samples. Additionally, it is thought that features can be extracted from the sample and its semantic neighbor samples to enhance text features. Furthermore, we enhance the original similarity matrix by incorporating related information. This step aims to improve the accuracy of similarity estimation by considering the enriched text features obtained in the previous step. Finally, we introduce an enhanced attention fusion mechanism. This mechanism adaptively fuses the similarity matrices from different modalities, creating a unified inter-modal similarity matrix. This fused matrix guides the learning of hash functions by preserving the most relevant information from each modality. Through comprehensive experiments on the three popular datasets, the suggested ESAFH method is thoroughly assessed. The findings show that on these datasets, ESAFH performs satisfactorily in cross-modal retrieval tasks. In conclusion, by boosting text features, improving the similarity matrix, and utilizing an attention fusion mechanism, ESAFH solves the shortcomings of current methods.

Based on traditional content-based image retrieval, the novel retrieval approach of using textual information to complement query images to retrieve target images is called text-image combination retrieval, which has attracted considerable attention in recent years. However, existing combination retrieval methods based on deterministic embedding cannot accurately measure the uncertainty of data. In this paper, we propose a CLIP-based Probabilistic Combination Retrieval model (CLIP-ProbCR), which utilizes probabilistic embedding and combining methods to add probability distributions to features after CLIP encoding through Gaussian probability density function multiplication rule for feature fusion. This method can quantify the uncertainty of instance embedding, simplifies the combined retrieval model architecture without complex auxiliary information with multi-level features or large architectures, and is validated on the Fashion IQ dataset and CIRR dataset, the experimental results demonstrate the effectiveness of the proposed model.

Despite the great success of existing cross-modal retrieval methods, existing unsupervised cross-modal hashing methods still suffer from common problems. First, the features extracted from the text are too sparse. Second, the similarity matrices of each different modality cannot be fused adaptively. In this paper, we propose Deep Enhanced-Similarity Attention Hashing (DESAH) to alleviate the above problems. Firstly, we construct a text encoder expanding graph convolutional neural network to simultaneously extract features of samples and their semantic neighbors to enrich text features. Secondly, we propose an enhanced attention fusion mechanism. The mechanism is used to adaptively fuse the similarity matrices within different modalities to form a unified inter-modal similarity matrix to guide the learning of hash functions. Extensive experiments have demonstrated that DESAH provides significant improvements in cross-modal retrieval tasks compared to baseline methods.

Cross-modal hashing is an important direction for multimodal data management and applications, which has recently received more and more attention. Unsupervised cross-modal retrieval does not rely on tag information and is more applicable to the real world. However it still faces some problems. Existing methods mainly encode for local features or global features. Due to the effect of negative samples, it is easy to cause noise interference. To solve these problems, we propose a Multi-view–enhanced modal fusion hashing for Unsupervised cross-modal retrieval (MUCH) to improve these problems. Firstly, we propose a multi-view network. Images inherently contain richer semantics, and we employ a multi-view network to observe the image from different perspectives and obtain the overall and local features of the image. Secondly, we introduce a noise cancellation module to approximate the cross-modal data feature alignment from both intra-modal and cross-modal perspectives before generating the hash code. Finally, we construct a distribution-based similarity weighting matrix to replace the graphical similarity matrix. And we performed multi-view enhancement experiments on JDSH and CIRH, with 1% to 2% enhancement over DAEH on all three datasets.

Due to the advantages of low storage cost and fast retrieval efficiency, deep hashing methods are widely used in cross‐modal retrieval. Images are usually accompanied by corresponding text descriptions rather than labels. Therefore, unsupervised methods have been widely concerned. However, due to the modal divide and semantic differences, existing unsupervised methods cannot adequately bridge the modal differences, leading to suboptimal retrieval results. In this paper, we propose CLIP‐based cycle alignment hashing for unsupervised vision‐text retrieval (CCAH), which aims to exploit the semantic link between the original features of modalities and the reconstructed features. Firstly, we design a modal cyclic interaction method that aligns semantically within intramodality, where one modal feature reconstructs another modal feature, thus taking full account of the semantic similarity between intramodal and intermodal relationships. Secondly, introducing GAT into cross‐modal retrieval tasks. We consider the influence of text neighbour nodes and add attention mechanisms to capture the global features of text modalities. Thirdly, Fine‐grained extraction of image features using the CLIP visual coder. Finally, hash encoding is learned through hash functions. The experiments demonstrate on three widely used datasets that our proposed CCAH achieves satisfactory results in total retrieval accuracy. Our code can be found at: https://github.com/CQYIO/CCAH.git .

Recently, cross-modal hashing has attracted much attention due to its low storage cost and fast query speed. Mean Average Precision (MAP) is the most widely used performance measure for cross-modal hashing. However, we found that the MAP scores do not fully reflect the quality of the top-K results for cross-modal retrieval because it neglects multi-label information and overlooks the label semantic hierarchy. In view of this, we propose a new performance measure named Normalized Weighted Discounted Cumulative Gains (NWDCG) by extending Normalized Discounted Cumulative Gains (NDCG) using co-occurrence probability matrix. To verify the effectiveness of NWDCG, we conduct extensive experiments using three popular cross-modal hashing schemes over two publically available datasets.

Cross-modal hashing (CMH) maps heterogeneous multiple modality data into compact binary code to achieve fast and flexible retrieval across different modalities, especially in large-scale retrieval. As the data don't need a lot of manual annotation, unsupervised cross-modal hashing has a wider application prospect than supervised method. However, the existing unsupervised methods are difficult to achieve satisfactory performance due to the lack of credible supervisory information. To solve this problem, inspired by knowledge distillation, we propose a novel unsupervised Knowledge Distillation Cross-Modal Hashing method (KDCMH), which can use similarity information distilled from unsupervised method to guide supervised method. Specifically, firstly, the teacher model adopted an unsupervised distribution-based similarity hashing method, which can construct a modal fusion similarity matrix.Secondly, under the supervision of teacher model distillation information, student model can generate more discriminative hash codes. In two public datasets NUS-WIDE and MIRFLICKR-25K, extensive experiments have proved the significant improvement of KDCMH on several representative unsupervised cross-modal hashing methods.