Abstract: Abstract: To address the issue of low sensitivity of existing few-shot object detection models to the feature parameters of new categories, which leads to unclear feature boundaries and category confusion, we propose a few-shot object detection algorithm based on new category feature enhancement and metric mechanisms. Firstly, a Cross-Domain Parameter Perception Module (CDPM) is proposed to improve the neck network, reconstructing the re-weighting operation of channel and spatial features. It combines dilated convolution with cross-stage information transfer and feature fusion to provide rich gradient information guidance and enhance the sensitivity of new category parameters. Meanwhile, a Integrated Correlated Multi-Feature Module (ICMF) is constructed before ROI Pooling to establish the correlation between features and dynamically optimize the fusion method of relevant features, enhancing significant features. The introduction of CDPM and ICMF effectively enhances the feature representation of new categories to alleviate feature boundary ambiguity. Additionally, to further reduce category confusion, an orthogonal loss function based on metric mechanisms (CohSep Loss) is proposed in the detection head to achieve intra-class feature aggregation and inter-class feature separation by measuring feature vector similarity. On the PASCAL VOC dataset, the proposed algorithm improves the average nAP50 by 5.3% compared to the baseline TFA. On the COCO dataset, the nAP for 10-shot and 30-shot settings is improved by 3.6% and 6.2%, respectively. Experimental results demonstrate that the proposed model achieves higher accuracy in few-shot object detection.

Key words: Keywords: transfer learning, few-shot object detection, feature enhancement, category confusion, orthogonal loss

摘要： 摘 要: 针对现有小样本目标检测算法中模型对新类别的特征参数敏感度较低，难以准确区分类别相关和类别无关参数，导致特征边界模糊，造成类别混淆的问题，提出了一种结合新类特征增强与度量机制的小样本目标检测算法。首先，提出跨域参数感知模块(CDPM)改进颈部网络，重构通道和空间的特征重加权操作，结合空洞卷积采用跨阶段的信息传递与特征融合方式，提供丰富的梯度信息导向并提升新类别参数敏感性。同时，在ROI Pooling前构造多元相关特征融合模块(ICMF)，建立特征之间的相关性并动态优化相关特征的融合方式，增强显著特征。CDPM与ICMF的引入有效增强了新类别的特征表示以减轻特征边界模糊。此外，为进一步减轻类别混淆，在检测头部分提出基于度量机制的正交损失函数(CohSep Loss)，通过度量特征向量相似度，实现类内特征聚合，类间特征分离。在PASCAL VOC数据集上，本算法相比基线TFA在nAP50的平均值上提升了5.3%；在COCO数据集中，10shot和30shot对应的nAP分别提升了3.6%和6.2%。实验结果表明，所提出的模型实现了更高精度的小样本目标检测。

关键词: 关键词: 迁移学习, 小样本目标检测, 特征增强, 类别混淆, 正交损失