Abstract:

To address the issues of low sensitivity to the feature parameters of new categories and difficulty in distinguishing category related and category unrelated parameters accurately of the existing few-shot object detection models， leading to unclear feature boundaries and category confusion， a Few-Shot Object Detection algorithm based on new categories Feature Enhancement and Metric Mechanism （FEMM-FSOD） was proposed. Firstly， a Cross-Domain Parameter perception Module （CDPM） was introduced to improve the neck network， thereby reconstructing re-weighting operations of channel and spatial features， and dilated convolution was combined with cross-stage information transfer and feature fusion to provide rich gradient information guidance and enhance the sensitivity of new category parameters. Meanwhile， an Integrated Correlated Multi-Feature module （ICMF） was constructed before Region of Interest Pooling （RoI Pooling） to establish correlation between features and optimize the fusion method of relevant features dynamically， thereby enhancing salient features. The introduction of CDPM and ICMF enhanced the feature representation of new categories effectively， so as to alleviate feature boundary ambiguity. Additionally， to further reduce category confusion， an orthogonal loss function based on metric mechanism， Coherence-Separation Loss （CohSep Loss）， was proposed in the detection head to achieve intra-class feature aggregation and inter-class feature separation by measuring feature vector similarity. Experimental results show that compared to the baseline algorithm TFA （Two-stage Fine-tuning Approach）， on PASCAL VOC dataset， the proposed algorithm improves the mAP50 （mean Average Precision （mAP） of new categories with threshold of 0.50） of 15 types of few-shot instance numbers by 5.3 percentage points； on COCO dataset， the proposed algorithm improves the mAP （mAP of new categories with threshold from 0.50 to 0.95） for 10shot and 30shot settings by 3.6 and 5.2 percentage points， respectively， realizing higher accuracy in few-shot object detection.

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

摘要：

针对现有的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和难以准确区分类别相关和类别无关参数，导致特征边界模糊以及类别混淆的问题，提出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（FEMM-FSOD）。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，并结合空洞卷积采用跨阶段的信息传递与特征融合方式，以提供丰富的梯度信息导向并提升新类别参数的敏感性；同时，在感兴趣区域池化（RoI Pooling）前构造多元相关特征融合模块（ICMF），以建立特征之间的相关性并动态优化相关特征的融合方式，从而增强显著特征。CDPM与ICMF的引入有效了增强新类别的特征表示，从而减轻特征边界模糊的现象。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（Coherence-Separation Loss），以通过度量特征向量相似度实现类内特征聚合和类间特征分离。实验结果表明，相较于基线算法TFA（Two-stage Fine-tuning Approach），在PASCAL VOC数据集上，所提算法在15种小样本实例个数的mAP50（阈值为0.50时新类别的平均精度均值（mAP））上提升了5.3个百分点；在COCO数据集上，所提算法在10shot和30shot对应的mAP（阈值为0.50~0.95时新类别的mAP）上分别提升了3.6和5.2个百分点，实现了更高精度的小样本目标检测。

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