Attention-enhanced vision transformer hashing for hybrid image retrieval

Large-scale image retrieval requires compact representations without substantially sacrificing retrieval accuracy. However, Vision Transformer Hashing (VTS) concatenates all output tokens before hash projection, resulting in a high-dimensional hashing head with considerable model and memory overhead. We replace this token-concatenation operation with learnable-query multi-head attention pooling, which aggregates ViT tokens into a compact, image-adaptive representation for supervised hash-code learning. We further evaluate the proposed model using a standard two-stage retrieval procedure. In Stage 1, the attention-pooled representation is mapped to a binary code for efficient Hamming-space candidate selection. In Stage 2, the final-layer CLS descriptor from the same shared ViT-B/16 is used to re-rank the shortlisted candidates by cosine similarity. Controlled experiments on MS-COCO and NUS-WIDE compare attention pooling with GeM, average, CLS, and VTS-style concatenation. The proposed hybrid configuration achieves 91.13% and 88.63% mAP@5000 on MS-COCO and NUS-WIDE, respectively, demonstrating performance comparable to the VTS-style Concatenation baseline under the same controlled configuration, with marginal gains of 0.15 and 0.31 percentage points. It also exceeds separately trained 32-bit attention hash-only models by 2.23 and 1.30 points. Relative to concatenation, attention pooling reduces model footprint and batch-1 GPU memory by 63.1% and 62.0%, respectively, while increasing throughput by 21.7%. These results show that attention pooling substantially improves the efficiency of VTS while preserving retrieval effectiveness, whereas continuous re-ranking provides most of the gain over Hamming-only retrieval.