Abstract: Abstract: Enriching the semantic information available in the support sets and query sets helps the model in few-shot scenarios to understand the target categories and then effectively complete the target detection task. Therefore, a model based on spatial context and pixel relationship is proposed to address the problem of insufficient feature information for model training in few-shot object detection. The spatial context module is used to assist the pixels in constructing a local context region, obtaining the semantics of the pixels in the region for the center pixel, and enriching the image feature information. In addition, to address the problem that spatial context easily introduces noisy information, the pixel context relation module is designed to utilize the inherent feature knowledge in the image to explore the relationship between pixels, and construct intra and inter-class relation maps to correct the defect that spatial context module easily introduces noisy information. Results from extensive experiments on the PASCAL VOC datasets and the more challenging MS COCO datasets show that the three ways of segmentation utilizing the VOC datasets resulted in improvements of 2.7, 2.0, and 1.3 percentage points for the 1-shot setting where samples are extremely sparse, and an average of 0.8 percentage points for all settings with 15 evaluation metrics. The average precision (AP) improved by 0.4 and 0.6 percentage points under the COCO datasets 10shot and 30shot, respectively, and the 50% intersection over union average precision improved by 9.8 and 8.7 percentage points, respectively. Compared with other few-shot object detection methods based on meta-learning or migration learning, the proposed model shows significant detection performance improvement under different task settings, demonstrating the effectiveness of the proposed modeling approach。

Key words: Keywords: few-shot object detection, insufficient feature information, spatial context, pixel relationship mapping

摘要： 摘 要: 丰富支持样本和查询样本中可用的语义信息，有助于小样本场景下模型理解目标类别，进而有效地实现目标检测任务。因此，针对小样本目标检测中模型训练特征信息不足的问题，提出一个基于空间上下文和像素关系的模型。利用空间上下文模块辅助像素构建局部上下文区域，为中心像素获取区域内像素语义，丰富图像特征信息。此外，针对空间上下文容易引入噪声信息的问题，设计了像素上下文关系模块利用图像中的原始特征知识探索像素之间的关系，构建类内和类间关系映射图，以纠正空间上下文模块容易引入噪声信息的缺陷。通过在PASCAL VOC数据集和更有挑战性的MS COCO数据集上进行的大量实验结果表明，利用VOC数据集进行三种方式划分时，样本极其稀缺的1shot设置下分别提升了2.7、2.0和1.3个百分点，在所有设置15个评价指标下平均提升了0.8个百分点。在COCO数据集10shot和30shot下，平均精度（AP）分别提升了0.4和0.6个百分点，50%的交并比平均精度分别提升了9.8和8.7个百分点。与其它基于元学习或迁移学习的小样本目标检测方法相比，所提出的模型在不同任务设置下有明显的检测性能提升，证明了模型方法的有效性。

关键词: 关键词: 小样本目标检测, 特征信息不足, 空间上下文, 像素关系映射