<![CDATA[This study evaluates a controlled content-based remote sensing image retrieval pipeline that combines frozen vision foundation models with binary hashing. Across five stratified seeds on PatternNet and NWPU-RESISC45, calibrated DINOv2–CLIP score fusion achieved mAP values of 0.8143 ± 0.0007 and 0.5693 ± 0.0020, respectively, improving over the stronger single backbone by 0.0582 and 0.0514. The two similarity spaces were related but not redundant, with Spearman correlations of ρ = 0.510 and 0.515 across 200,000 sampled pairs. The contribution of texture features was dataset- and backbone-dependent: DINOv2–texture received a weight of 0.1 on PatternNet, whereas CLIP–texture and three-stream fusion assigned zero weight to texture; forcing texture into a 32-bit representation reduced mAP by 0.0147 and 0.0510. Among the unsupervised hashing methods, 64-bit iterative quantization (ITQ) exceeded the full-precision fused baseline by 0.0533 and 0.0445 mAP on PatternNet and NWPU-RESISC45, respectively. The corresponding paired parametric tests were significant, although the exact five-pair sign-flip test remained resolution-limited (p = 0.0625). A supervised CSQ-style head achieved mAP values of 0.9886 and 0.9087 at 64 bits; however, this improvement cannot be attributed to compression alone because gallery labels were used during training. In a controlled 200,000-vector benchmark constructed through deterministic repetition, 64-bit flat Hamming search required 1.6 MB and 0.674 ms/query, compared with 1.02 GB and 5.343 ms/query for 1,280-dimensional float vectors. Overall, the results support DINOv2–CLIP fusion and 64–128-bit ITQ as practical label-free choices, while restricting the scalability claim to the tested flat-search setting.]]>
Copyrights © 2026
