Remote Sensing Cross-Modal Text-Image Retrieval Using Multi-Grade Dynamic Feature Fusion

Recent advances in technology have led to a significant increase in the size and number of remote sensing image datasets, presenting challenges in efficient data retrieval. This is particularly difficult because of the heterogeneity of the data and the large semantic gap between the text and image modalities. Traditional retrieval techniques are not well-suited to handle multimodal remote sensing data. To address this, the paper introduces a novel framework designed for the retrieval of images based on text queries and vice versa. Existing remote sensing text-image retrieval methods often focus on high-level features derived from RS images, neglecting valuable low-level features that provide insights into target relationships and saliency. The proposed model emphasizes vision feature extraction and integrates both macro- and micro-vision features, enhancing each through the other. Deep learning techniques are employed to generate comprehensive text and image representations. Once these embeddings are generated, their similarity is computed, and a re-ranking algorithm is applied, utilizing the k-nearest neighbors from the retrieval outcomes to improve accuracy. The model is evaluated using the recall metric, showing superior performance compared to conventional methods on the RSITMD and RSICD datasets.