<p>This paper introduces RGFRCap, an innovative image captioning framework that leverages retrieval-guided semantic feature refinement to enhance caption quality. At its core, RGFRCap integrates an image-text retrieval (ITR) module to fetch candidate captions that closely match the input image, serving as conditional priors for subsequent processing. These priors guide a semantic feature filtering (SFF) mechanism, which refines semantic information extracted via object detection and semantic segmentation, focusing on relevant objects, attributes, and pixel-level details. The refined semantic features are then amalgamated with region-specific visual features in a visual-semantic fusion (VSF) module, enriching the visual representation. A vision-language transformer decoder utilizes this enhanced representation to produce precise and contextually rich captions. Empirical evaluations on MSCOCO, Flickr30K, and two custom traffic datasets (City_cap and FoggyCity_cap) showcase RGFRCap’s superior captioning performance, surpassing existing methods on several benchmarks. The codebase and datasets are publicly accessible at <a href="https://github.com/fjq-tongji/RGFRCap,">https://github.com/fjq-tongji/RGFRCap,</a> fostering further research in the field.</p>