An Adaptive Fusion Network Integrating Multi-Source Feature for Remote Sensing Image-Text Retrieval

The rapid evolution of deep learning technologies, particularly the widespread adoption and improvement of deep neural networks, Transformer architectures, and image-text contrastive learning, has driven significant advancements in the domain of cross-modal retrieval. However, the performance of these retrieval models are limited by the complex and variable nature of remote sensing images. In this paper, we proposed a specialized image-text retrieval network for remote sensing images. Our highlight is the proposal of an adaptive feature fusion network and a threshold filtering module in the image feature extraction section. The adaptive feature fusion network learns feature map weights to dynamically combine global features from the Transformer and local details from the CNN, effectively mitigating multi-scale and small object issues in remote sensing imagery. Additionally, the flexible threshold filtering module leverages an attention mechanism to suppress redundant features, further improving the semantic representation of small objects. Experimental validation on two public datasets, RSITMD and RSICD, demonstrates that this method not only accurately achieves cross-modal retrieval between remote sensing images and text but also outperforms existing baseline methods, exhibiting superior performance.