2.1 Single-tower models

Single-tower (single-stream) architecture models process image and text features through a shared encoder, like a transformer, as shown in Figure 2 (a). These models usually combine the two input modalities early in the network and jointly process them through shared encoders. The main motivation behind single-tower models is their ability to directly learn joint representations of the two modalities, capturing complex interactions between them. By using shared layers to process both modalities together, these models aim to learn rich, fused representations that benefit cross-modal retrieval tasks.

In this review, we focus on the role of modality fusion in enhancing retrieval performance. Particularly, single-tower models exemplify early fusion by embedding image and text inputs into a unified semantic space through shared encoders. This strategy enables deeper interactions between modalities, yielding richer representations. Additionally, MLLM-based approaches, such as BLIP-2, implement hybrid fusion through modules like Q-Former, allowing semantic alignment across modalities at multiple stages.

The ViLT (Vision and Language Transformer) model [8] presents an innovative method for multi-modal training, drawing inspiration from the Vision Transformer (ViT) mechanism. In contrast to earlier methods that needed an object detector for region-level feature extraction, ViLT directly splits images into patches, performs linear embedding, and uses these as transformer inputs. Text data is also embedded and merged with image embeddings for joint training, significantly enhancing learning and inference efficiency. ViLT employs three pre-training objectives: Image-Text Matching (ITM), Masked Language Modeling (MLM), and Word Patch Alignment (WPA). For fine-tuning in cross-modal retrieval, ViLT initializes the similarity score head from the pre-trained ITM head and fine-tunes it with cross-entropy loss to maximize positive pair scores. Experimental results indicate that ViLT drastically reduces per-instance processing time from 900 milliseconds to 15 milliseconds, showcasing its efficiency and innovation in multi-modal learning.

Traditional pre-trained models for computer vision (CV) and natural language processing (NLP) perform well independently but face challenges with cross-modal tasks involving lengthy natural language inputs and intricate visual elements. Unicoder-VL [9] utilizes a multi-layer Transformer to learn joint representations of vision and language via cross-modal pre-training. It uses three tasks: MLM, Masked Object Classification (MOC), and Visual-linguistic Matching (VLM) The model processes linguistic and visual content simultaneously, effectively learning context-aware representations and predicting relationships between images and texts. Pre-training on large-scale image-caption pairs allows it to excel in downstream tasks such as image-text retrieval and visual commonsense reasoning. Unicoder-VL achieves state-of-the-art results in image-text retrieval on the MSCOCO and Flickr30K datasets, showcasing strong generalization abilities. However, its reliance on pre-training datasets might limit performance on tasks that require domain-specific knowledge.

A flexible model is needed to handle various vision-and-language tasks, capturing detailed semantics from both modalities without complex architectures. VisualBERT [10] integrates BERT with pre-trained object detection systems, processing image features and text together using Transformer layers. It is pre-trained on the COCO dataset using visually-grounded language model objectives such as masked language modeling and sentence-image prediction The model's design enables it to implicitly align language elements and image regions through self-attention, capturing intricate associations without explicit supervision. VisualBERT's design emphasizes simplicity and flexibility in handling diverse tasks.

Single-stream methods represent a powerful approach for cross-modal retrieval, leveraging unified Transformer architectures to effectively bridge the gap between different modalities. While these models perform well on general datasets, fine-tuning them for specific domains may require additional data and computational adjustments.

2.2 Dual-tower models

Dual-stream cross-modal methods, aim to integrate and process information from multiple modalities, such as text, images, and audio. These methods are characterized by their ability to handle the heterogeneity and complexity inherent in multimodal data, thereby facilitating a richer and more comprehensive understanding and generation of content. The primary challenge addressed by dual-stream cross-modal methods is the effective alignment and fusion of disparate data types, which often possess different structures, noise levels, and contextual nuances. The dual stream cross-modal approach typically involves two parallel processing streams, each dedicated to handling a specific modality, as shown in Figure 2 (b).

ViLBERT [11] aims to tackle the challenge of jointly understanding and reasoning about vision and language, which is difficult due to the inherent differences and complexities of each modality. It employs a two-stream model in which one stream processes visual information and the other processes linguistic information. These streams interact via a co-attentional Transformer layer that enables each modality to attend to the other. The key innovation is the co-attentional Transformer layer, which facilitates the interaction between visual and linguistic representations, allowing the model to learn rich, joint representations of both modalities.

CLIP [12] meets the need for models that can understand and connect images and text flexibly, particularly for zero-shot learning tasks where the model must generalize to new concepts without explicit training. CLIP employs separate encoders for images and text, training them with a contrastive loss to align image and text embeddings in a shared space. The model is trained on a vast dataset of images and their corresponding captions from the internet. The key innovation is using contrastive learning to align visual and textual representations, enabling the model to perform zero-shot learning by leveraging the rich, diverse data it was trained on. CLIP demonstrates impressive performance on various tasks without fine-tuning, including image classification, image-text retrieval, object detection, and generating text descriptions for images.

In [14], the authors introduce ALIGN (A Large-scale ImaGe and Noisy-text embedding), which utilizes a massive dataset of over one billion image-alt text pairs collected from the web with minimal filtering. The core of ALIGN is a straightforward dual-encoder architecture that employs contrastive learning to align visual and language representations in a shared embedding space. The ALIGN model uses a dual-encoder architecture with separate encoders for images and text. The encoders are trained with a contrastive loss to align the embeddings of matching image-text pairs. During training, the model applies simple frequency-based filtering on the dataset. The contrastive loss function helps in bringing together the embeddings of matched pairs and separating those of non-matched pairs. ALIGN achieves $76.4\%$ top-1 accuracy on ImageNet without using any of its training samples and sets new state-of-the-art results on Flickr30K and MSCOCO benchmarks. In addition to the basic dual-encoder designs, some recent studies further enhance retrieval quality by promoting representation diversity. For instance, Kim et al. [13] proposed a method that integrates a set of diverse embeddings to enrich the semantic space, improving the robustness of cross-modal retrieval across varying query intents and data distributions.

Dual-stream methods offer a robust framework for cross-modal retrieval by utilizing specialized pathways for different modalities and aligning their outputs in a shared space. By effectively aligning embeddings and using tailored processing, these models achieve strong performance in retrieving relevant content across heterogeneous data types, showcasing their value in multimodal applications.

2.3 Real-value representation models

Non-hashing methods based on real-valued representations effectively reduce the semantic gap between different modalities by learning dense feature representations, thereby enhancing retrieval precision. By employing deep learning methods to model features of various modalities and extract deep semantic features, these methods effectively address the issue of feature heterogeneity in cross-modal data. They also emphasize semantic correspondence between modalities, narrowing the semantic gap to improve the accuracy of cross-modal data matching, thereby increasing retrieval precision.

ACMR [15] tackles the challenge of aligning visual and textual data for cross-modal retrieval tasks, where traditional methods often fail to bridge the semantic gap between different modalities. The proposed solution involves employing adversarial training to learn robust cross-modal representations. Specifically, ACMR utilizes a dual-stream architecture where each modality is processed separately, with an adversarial loss to align the embeddings in a shared space. The key innovation of ACMR is the integration of adversarial learning, which encourages the model to produce modality-invariant features. This approach ensures that visual and textual representations are more closely aligned, thereby improving retrieval accuracy. ACMR significantly enhances the performance of cross-modal retrieval tasks, demonstrating improved alignment between visual and textual data and higher retrieval accuracy compared to non-adversarial methods. However, adversarial training can be complex and computationally intensive, and it may lead to potential instability during training.

DSCMR [16] addresses the challenge of learning effective representations for cross-modal retrieval tasks, where existing methods often struggle to capture the complex relationships between different modalities. The proposed solution employs a deep supervised approach that utilizes labeled data to learn discriminative features for each modality. DSCMR uses a dual-stream network with deep neural networks for both visual and textual data, supervised by a cross-modal ranking loss. The innovation in DSCMR lies in its application of deep supervision and a cross-modal ranking loss, ensuring that the learned representations are both discriminative and aligned across modalities. DSCMR achieves state-of-the-art performance in cross-modal retrieval tasks, showcasing the effectiveness of deep supervision and ranking-based training objectives in improving retrieval accuracy. However, DSCMR requires large amounts of labeled data and is potentially prone to overfitting to specific datasets.

IEFT [17] tackles the challenge of enhancing feature interactions for cross-modal retrieval, where traditional models often fail to fully capture the intricate relationships between visual and textual data. The proposed solution, Interacting-Enhancing Feature Transformer (IEFT), uses a Transformer-based architecture to enhance feature interactions between modalities. IEFT processes visual and textual features in separate streams and employs attention mechanisms to integrate them. The key innovation of IEFT is its use of Transformer-based attention mechanisms to enhance interactions between visual and textual features, allowing the model to learn richer and more nuanced representations. IEFT demonstrates superior performance on cross-modal retrieval benchmarks, benefiting from enhanced feature interactions and the powerful representation capabilities of Transformers.

COTS [18] addresses the difficulty of effectively combining visual and textual information for cross-modal retrieval, where existing methods may not fully leverage the potential of collaborative learning between modalities. The solution involves a Collaborative Two-Stream (COTS) architecture, where two streams process visual and textual data independently but collaborate through shared intermediate representations and alignment losses. The innovation in COTS lies in its collaborative learning mechanism, which ensures that the two streams not only process their respective modalities effectively but also learn from each other through shared representations. While collaborative learning enhances feature alignment and robust performance across various tasks, it increases complexity due to collaboration mechanisms and potential synchronization issues between streams.

TEAM [19] addresses the issue of aligning token embeddings from different modalities for cross-modal retrieval, where conventional methods may not fully capture the semantic relationships between visual and textual data. The proposed solution, Token Embeddings AlignMent (TEAM), employs alignment strategies to ensure that token embeddings from different modalities are closely related in a shared space. TEAM utilizes dual-stream networks with alignment losses to achieve this goal. TEAM's key innovation is its specific focus on token-level alignment, ensuring that individual tokens from text and corresponding visual elements are accurately aligned in the embedding space. TEAM significantly improves cross-modal retrieval performance by ensuring precise alignment of token embeddings, leading to better semantic understanding and retrieval accuracy. However, it incurs potentially high computational costs for fine-grained alignment and complexity in managing token-level interactions.

2.4 Binary representation models

Real-valued cross-modal image-text retrieval methods based on deep learning use feature vectors directly obtained from feature extraction for modeling and retrieval. However, with the explosive growth of multimedia data, such as short videos on TikTok or image-text information on Weibo, multimodal data often reaches hundreds of thousands, millions, or even billions of instances. This requires that the retrieval process for multimodal data ensures both precision and efficiency. Among various retrieval methods, hashing methods have gained widespread attention due to their low storage cost, efficiency, and fast retrieval speed, making them more suitable for large-scale datasets.

Hashing methods map feature vectors from the original feature space to binary codes (Hamming space) to save storage space and increase retrieval speed while maintaining the similarity between data points during the mapping process. Subsequently, the Hamming distance between the hash codes of the query data and those in the database is calculated for similarity ranking, ultimately yielding the retrieval results. Calculating the Hamming distance is faster than other distance metrics such as Euclidean and cosine distances. Additionally, storing data as binary codes rather than real-valued ones reduces the storage requirements for retrieval tasks.

Learning hash functions mainly involves dimensionality reduction and quantization. Dimensionality reduction maps the information from the original space to a lower-dimensional space, such as mapping an image's original pixel space information to a lower-dimensional (e.g., tens of dimensions) representation. Quantization involves linear or nonlinear transformations of the original features and binary segmentation of the feature space to produce hash codes. As mentioned in the problem definition section of cross-modal retrieval, there is a semantic gap between different forms (modalities) of data representation. Minimizing this semantic gap remains a primary challenge for cross-modal retrieval hashing methods. Generally, there are two approaches to address this: one is learning a unified hash code, and the other is using supervised information, such as labels, to collaboratively represent and minimize the distance between hash codes of semantically relevant instances.

DCMH [20] addresses the challenge of efficiently retrieving relevant data across different modalities (e.g., text and images) by using hashing techniques to map high-dimensional data into compact binary codes. The proposed solution utilizes a deep learning framework to generate hash codes for each modality through learning shared representations. These representations are optimized to maintain semantic similarity across different modalities, ensuring related items have similar hash codes. This is the first use of deep hashing neural networks to learn these representations, allowing the model to capture complex relationships between modalities and generate more accurate hash codes.

UDCMH [21] addresses the challenge of cross-modal retrieval without labeled data, which is significant since traditional supervised methods rely heavily on labeled training examples. The key innovation is the unsupervised learning approach, which eliminates the need for labeled data and still achieves effective cross-modal retrieval by learning from the data's inherent structure. This approach demonstrates strong performance in cross-modal retrieval tasks, especially in scenarios where labeled data is scarce or unavailable. However, its performance may not match supervised methods on well-labeled datasets and may be sensitive to the quality of the data structure.

SSAH [22] tackles the challenge of generating robust hash codes for cross-modal retrieval by leveraging the advantages of both self-supervised learning and adversarial training. Self-supervised learning generates initial hash codes, while adversarial training refines these codes to ensure they are modality-invariant and semantically meaningful. This combination enables the model to learn effective representations without the need for extensive labeled data. SSAH achieves enhanced retrieval performance and robustness, demonstrating the effectiveness of its novel training strategy.

Bi-CMR [23] is the first to recognize that the assumption "label annotations reliably reflect instance relevance" conflicts with human perception. It proposes a new evaluation method to guide the learning of instance hash codes consistent with human perception. Bi-CMR introduces a novel bidirectional reinforcement-guided hashing method that reinforces hash code learning through mutual promotion. The key innovation is using reinforcement learning to dynamically adjust and improve the hashing process, ensuring the generated hash codes are effective for cross-modal retrieval. Bi-CMR demonstrates superior performance in cross-modal retrieval tasks, with hash codes that are well-aligned and optimized for retrieval accuracy.

DCHMT [24] tackles the challenge of effectively integrating and hashing data from multiple modalities using a unified framework. It constructs a multi-modal transformer to capture detailed cross-modal semantic information and introduces a micro-hashing module to map modal representations into hash codes. UCMFH tackles the need for effective cross-modal retrieval without labeled data, focusing on learning robust hash codes through unsupervised methods. The proposed solution uses unsupervised contrastive learning to generate hash codes. By leveraging contrastive learning, the model maximizes the similarity between related items across modalities while minimizing the similarity between unrelated items. UCMFH demonstrates strong performance in unsupervised cross-modal retrieval tasks, achieving high accuracy and robustness by effectively learning from the inherent structure of data.

Overall, real-valued representations are suitable for tasks that require high precision, while hashing representations are ideal for applications that need rapid, large-scale retrieval.

4.1 Datasets

The researchers have proposed various datasets for cross-modal image-text retrieval, including Wikipedia [25], NUS-WIDE [26], TC-12 [27], Flickr [28], Pascal Sentence [29], etc. The most frequently used datasets are summarized as MSCOCO [30] and Flickr30K [28]. MS COCO dataset contains $123,287$ images from the Microsoft Common Objects in Context (COCO) dataset, each paired with five human-generated textual captions. After removing rare words, the average caption length is $8.7$ words. The dataset is divided into $82,783$ training image-text pairs, $5,000$ validation pairs, and $5,000$ test pairs. Model evaluations are conducted on five folds of $1,000$ test pairs and the entire set of $5,000$ test pairs. Flickr30K0 comprising $31,000$ images sourced from the Flickr website, each image in this dataset is annotated with five textual descriptions. The dataset is split into three sections: $1,000$ image-text pairs for validation, $1,000$ pairs for testing, and the remaining for training.

4.2 Evaluation

We summarize the following evaluation metrics widely used to assess cross-modal retrieval tasks: Mean Average Precision@K (MAP@K): MAP calculates the average precision for each query and then averages these values over all queries. In the experimental validation of MLLMs, the R@n metric is commonly used, indicating the proportion of queries for which at least one correct result is retrieved within the top-n results.

4.3 Result Analysis

In this section, we present the accuracy of several representative methods in cross-modal retrieval tasks. As shown in Tables 2- 4, we compare the accuracy of cross-modal retrieval methods using common measures for each task. Based on the presented performance, we can summarize the following observations: