[Paper Review] Visual Semantic Reasoning for Image-Text Matching

Kunpeng Li, Yulun Zhang+3 more|arXiv (Cornell University)|Sep 6, 2019

Multimodal Machine Learning Applications40 references21 citations

TL;DR

This paper proposes Visual Semantic Reasoning Network (VSRN), a novel model that enhances image-text matching by performing region-based semantic reasoning via Graph Convolutional Networks (GCNs) and global semantic reasoning using a gate and memory mechanism. The method generates interpretable, high-level visual representations that capture key objects and relationships, achieving state-of-the-art performance on MS-COCO and Flickr30K with relative improvements of 6.8% and 12.6% on image retrieval, respectively.

ABSTRACT

Image-text matching has been a hot research topic bridging the vision and language areas. It remains challenging because the current representation of image usually lacks global semantic concepts as in its corresponding text caption. To address this issue, we propose a simple and interpretable reasoning model to generate visual representation that captures key objects and semantic concepts of a scene. Specifically, we first build up connections between image regions and perform reasoning with Graph Convolutional Networks to generate features with semantic relationships. Then, we propose to use the gate and memory mechanism to perform global semantic reasoning on these relationship-enhanced features, select the discriminative information and gradually generate the representation for the whole scene. Experiments validate that our method achieves a new state-of-the-art for the image-text matching on MS-COCO and Flickr30K datasets. It outperforms the current best method by 6.8% relatively for image retrieval and 4.8% relatively for caption retrieval on MS-COCO (Recall@1 using 1K test set). On Flickr30K, our model improves image retrieval by 12.6% relatively and caption retrieval by 5.8% relatively (Recall@1). Our code is available at https://github.com/KunpengLi1994/VSRN.

Motivation & Objective

• To address the visual-semantic discrepancy in image-text matching by enhancing image representations with high-level semantic concepts.
• To model both local region relationships and global semantic correlations in images for better alignment with text descriptions.
• To design a simple, interpretable reasoning mechanism that improves feature representation beyond local CNN features.
• To validate that the learned image representation captures key objects and semantic relationships through attention visualization.
• To achieve state-of-the-art performance on standard image-text matching benchmarks.

Proposed method

• The model uses Faster R-CNN to extract salient image regions as input for reasoning.
• It constructs a graph topology connecting image regions and applies Graph Convolutional Networks (GCNs) to propagate and enrich features with semantic relationships.
• A gate and memory mechanism is introduced to perform global semantic reasoning, selectively attending to discriminative features and gradually building a holistic image representation.
• The final image representation is computed by aggregating region features through a differentiable attention mechanism based on similarity to the global representation.
• The method visualizes attention scores between the final image representation and each region to interpret model behavior.
• The model is trained end-to-end with a contrastive ranking loss to optimize image-text matching performance.

Experimental results

Research questions

• RQ1 Can a reasoning mechanism that models relationships between image regions improve visual-semantic alignment in image-text matching?
• RQ2 Does global semantic reasoning using a gate and memory mechanism enhance the discriminative power of image representations?
• RQ3 Can the model's internal reasoning be interpreted to verify that it captures key semantic concepts from text captions?
• RQ4 Does the proposed reasoning-based representation outperform existing methods that rely on local features or pairwise attention?
• RQ5 How well does the model generalize to complex and cluttered scenes in standard benchmarks?

Key findings

• On MS-COCO, VSRN achieves a 6.8% relative improvement in image retrieval (Recall@1 on 1K test set) and a 4.8% relative improvement in caption retrieval over the previous state-of-the-art method.
• On Flickr30K, the model improves image retrieval by 12.6% relatively and caption retrieval by 5.8% relatively (Recall@1), demonstrating strong generalization.
• Qualitative results show that the model retrieves correct matches in top-3 for complex scenes, with attention maps highlighting key objects and semantic concepts.
• Attention visualization confirms that the model's final representation strongly correlates with regions containing key semantic concepts, validating interpretability.
• The model successfully distinguishes between visually similar images, such as differentiating a person cleaning a car from a person skiing based on semantic cues.
• The results indicate that enhanced global representation via reasoning is more effective than complex pairwise attention mechanisms for image-text matching.