Abstract
Few-shot object detection (FSOD) addresses the challenge of object recognition under limited annotation conditions, offering practical advantages for smart agriculture, where large-scale labeling of diverse fruit cultivars is often infeasible. To handle the visual complexity of orchard environments-such as occlusion, subtle morphological differences, and dense foliage-this study presents a lightweight tri-modal fusion framework. The model initially employs a CLIP-based semantic prompt encoder to extract category-aware cues, which guide the Segment Anything Model (SAM) in producing structure-preserving masks. These masks are then incorporated via a Semantic Fusion Module (SFM): a Mask-Saliency Adapter (MSA) and a Feature Enhancement Recomposer (FER), enabling spatially aligned and semantically enriched feature modulation. An Attention-Aware Weight Estimator (AWE) further optimizes the fusion by adaptively balancing semantic and visual streams using global saliency cues. The final predictions are subsequently generated by a YOLOv12 detection head. Experiments conducted on four fruit detection benchmarks-Cantaloupe.v2, Peach.v3, Watermelon.v2, and Orange.v8-demonstrate that the proposed method consistently surpasses five representative FSOD baselines. Performance improvements include +7.9% AP@0.5 on Cantaloupe.v2, +5.4% Precision on Peach.v3, +7.4% Precision on Watermelon.v2, and +5.9% AP@0.75 on Orange.v8. These results underscore the model's effectiveness in orchard-specific scenarios and its potential to facilitate cultivar identification, digital recordkeeping, and cost-efficient agricultural monitoring.