Visual grounding of remote sensing images with multi-dimensional semantic-guidance

Visual grounding in remote sensing images aims to accurately locate specified targets based on query expressions. Existing methods often use separate feature extractors to independently process visual and textual features. However, this approach results in initial features that lack correlation between the two modalities, hindering effective feature fusion and limiting localization precision. To address this challenge, we propose a novel framework called MSVG, which enhances visual grounding accuracy through a multidimensional text-image alignment module and a visual enhancement fusion module. The multi-dimensional text-image alignment module employs both channel-wise and spatial-wise alignment at various stages of visual feature extraction, guiding the generation of visual features in a manner that increases their relevance to the accompanying textual descriptions. Meanwhile, the visual enhancement fusion module refines feature relevance by learning contextual features, effectively excluding objects and backgrounds unrelated to the target. Experiments demonstrate that our approach achieves a remarkable accuracy of 83.61% on DIOR-RSVG dataset, representing a substantial 6.83% improvement over previous methods and setting a new benchmark.