Target reindexing and residual reconstruction enable reliable retrieval augmented few shot recognition

Retrieval-augmented few-shot recognition supplements sparsely annotated target classes with external visual samples, but open retrieval results may exhibit unstable quality, semantic mismatch, and cross-domain feature shifts. Directly using query-derived labels or treating retrieved images as ordinary training data can therefore distort class representations and decision boundaries. We propose the Target Re-indexing and Residual Reconstruction Network (TRRN), which incorporates external samples under target-task constraints. TRRN first constructs class anchors from the few-shot support set and estimates the support of each retrieved sample for every target class, thereby remapping external samples into the current class space without directly inheriting their retrieval labels. It then models the components unexplained by the target-anchor projection as residuals and combines class-support weights, confidence gating, and residual magnitudes to estimate reliable sample contributions. The retained information supplements the few-shot anchors, while the class-wise mean residual direction is suppressed during classifier reconstruction. External images are retrieved from LAION-400 M through class-name-based text-to-image retrieval using a frozen CLIP ViT-B/16 encoder, with 20 valid images retained per class after deduplication and overlap removal. Experiments on FGVCAircraft and DTD under full-way 4-shot, 8-shot, and 16-shot protocols show that TRRN achieves the highest Mean Accuracy in all six settings and outperforms the strongest retrieval-augmented baseline by 1.11–1.78% points. Paired tests yield $$p<0.001$$ , while ablation and sensitivity analyses further confirm the contributions of target re-indexing, confidence gating, and residual suppression.