Abstract: Existing few-shot object detection methods often have the problem of data distribution shift when amplifying samples, and the performance of classification tasks is easily affected by localization tasks. Aiming at the above problems, a new few-shot object detection algorithm is proposed based on the Faster R-CNN framework. The classification correction module (CCB) , sample amplification module (SAB) , and gradient control layer (GCL) are introduced to improve performance. CCB uses an offline strong classification network to correct the final results of the detector. SAB uses the base class information to modify the distribution of the new class samples in the feature domain, so as to complete the amplification of the new class samples by sampling from the modified distribution. In gradient backpropagation, the information of the base class and new class received by the backbone network are restricted by GCL. The experimental results on PASCAL VOC and COCO datasets show that, compared with the latest known algorithm results, the proposed few-shot object detection algorithm improves the detection effect when the number of samples is small. The maximum improvement can reach 5.1% on PASCAL VOC, a public dataset. It also reaches up to 1.9% improvement on the more difficult dataset COCO. Therefore, the proposed few-shot detection framework has good robustness and generalization ability at the same time.

Key words: few-shot learning, object detection, data amplification, gradient control

摘要： 现有小样本目标检测方法在扩增样本时往往存在数据分布偏移问题，同时分类任务性能容易受定位任务影响。针对上述问题，提出一种新的小样本目标检测算法。该算法在Faster R-CNN框架基础上引入分类校正模块（CCB）、样本扩增模块（SAB）和梯度限制层（GCL）改善性能。CCB使用离线的强分类网络对检测器最终结果进行校正；SAB在特征域利用基类样本信息修正新类样本分布，从而在修正的分布中进行采样完成新类样本扩增；在梯度反向传播中通过GCL限制主干网络接收的基类和新类信息。在PASCAL VOC和COCO数据集上的实验结果表明，相较于目前已知的最新算法结果，提出的小样本目标检测算法在样本数量很小的情况下提升了检测效果，在公共数据集PASCAL VOC上最高提升可以达到5.1%，更难的数据集COCO上最高提升可达到1.9%，同时拥有很好的鲁棒性和泛化能力。

关键词: 小样本学习, 目标检测, 数据扩增, 梯度限制