Abstract: Existing few-shot object detection (FSOD) algorithms often suffer from insufficient class-level prototype generation accuracy and loss of detail information, which limits the feature representation capacity in target regions. To address this issue, a method called Agent Prototypes Aggregation (APA) for FSOD is proposed. The proposed method utilizes agent attention to distill detail-rich prototypes from the support features, which are then accurately assigned to the query feature map based on their correlation, significantly enhancing the feature representation in target instance regions. Additionally, a Wavelet Convolution Enhancement Module (WCEM) and an Adaptive Multi-Relation Fusion Module (AMRF) are designed to optimize global feature extraction and high-level feature fusion, respectively. Experimental results demonstrate that, on the PASCAL VOC dataset with three novel class splits,APA-FSOD achieves a maximum nAP50 improvement of 1.1 percentage points compared to the baseline VFA (Variational Feature Aggregation). On the MS COCO dataset under the 30-shot setting, nAP is improved by 1 percentage point compared to the meta-learning method SMPCCNet (Support-query Mutual Promotion and Classification Correction Network), significantly enhancing the accuracy of few-shot object detection.

Key words: few-shot object detection, meta learning, agent prototypes distillation, wavelet convolution, multi-relation fusion

摘要： 针对现有小样本目标检测算法中类级原型生成精度不足、细节信息缺失导致目标区域特征表达能力受限的问题，提出了一种基于代理原型聚合（APA）的小样本目标检测算法（APA-FSOD），该算法通过代理注意力将支持特征蒸馏为细节丰富的原型，并基于相关性实现原型向量在查询特征图上的精准分配，显著强化了目标实例区域的特征表达能力。此外，还设计了小波卷积增强模块（WCEM）和自适应多关系融合模块（AMRF），分别用于优化算法的全局特征提取和高级特征融合。实验结果表明，在 PASCAL VOC 数据集的三种新类划分下，APA-FSOD 的 nAP50 相较于基线 VFA（Variational Feature Aggregation）最高提升了 1.1 个百分点。在 MS COCO 数据集的 30-shot 设置下，与元学习方法 SMPCCNet（Support-query Mutual Promotion and Classification Correction Network）相比，nAP 提升了 1 个百分点，显著提高了小样本目标检测的精度。

关键词: 小样本目标检测, 元学习, 代理原型蒸馏, 小波卷积, 多关系融合