FGS-CNCLIP: fine-grained street-view image geolocation model with CN-CLIP

Accurate image geolocation remains challenging, particularly for precise localization of street-view imagery in complex urban environments. This paper proposes a fine-grained street-view geolocation model, termed FGS-CNCLIP. The framework combines fine-grained geographic region partitioning, street-view feature fusion, and CLIP-based multimodal learning with enhanced discriminative capability. A region construction algorithm is designed to optimize geographic partitioning and sample classification, improving localization accuracy while maintaining computational efficiency. Geographic metadata are further incorporated to reduce the impact of visual homogenization in adjacent urban areas. CN-CLIP, a Chinese vision-language pretrained model, is adopted as the backbone to better support Chinese place semantics and local data characteristics. Experiments were conducted on more than 300,000 street-view images collected from Xihu District, Hangzhou, with additional evaluation on 4,903 open street-view images covering 312.43 km2. The proposed model achieves over 96% accuracy at the 500 m localization level within the trained city, demonstrating strong potential for region-specific fine-grained geolocation tasks.