基于可学习记忆特征金字塔网络的小样本目标检测

Few-shot Object Detection via Learnable Memory Feature Pyramid Network

现阶段,部分行业应用场景数据难以获取,从而产生的小样本问题成为制约深度学习技术应用推广的重要因素。本文通过小样本方法来提升模型在数据缺乏情况下的表现,降低深度学习模型对数据的依赖性,提出一种基于可学习记忆特征金字塔网络来保留更干净的多尺度特征信息用于分类器预测。借助自适应特征融合模块,让网络自行选择不同层级特征间的侧重比,最大化保留不同尺度的判别性特征信息。同时还加入回溯特征对齐模块,用于缓解特征层堆叠时引入的特征混淆效应。实验结果表明,通过克服样本依赖性可以有效地提升模型性能,改进后的模型可以在COCO数据集和VOC数据集上超越其他现有同类型的模型。特别地,在VOC数据集中将先验参数k设置为5的情况下,nAP50提高了4.8达到44.7;在COCO数据集中将先验参数k设置为30的情况下,nAP50提高了4.0达到29.4。

At present,it is difficult to obtain the data of some industry application scenarios,and the problem of few shot has become an important factor restricting the application and promotion of deep learning technology.In this paper,few shot method is adopted to improve the performance of the model in the absence of data and reduce the dependence of the deep learning model on data,and few-shot object detection via learnable memory feature pyramid network is proposed to retain cleaner multi-scale feature information for classifier prediction.With the help of the adaptive feature fusion module,the network can choose the emphasis ratio among the features of different levels to maximize the retention of discriminant feature information of different scales.At the same time,we also add a retrospective feature alignment module to alleviate the feature confusion effect introduced by stacking feature layers.The experimental results show that the model performance can be effectively improved by overcoming the dependence on data,and the improved model can surpass other existing models of the same type in the COCO dataset and VOC dataset.In particular,when the prior parameter k is set to 5 in VOC dataset,nAP50 increases by 4.8 to 44.7;when the prior parameter k is set to 30 in COCO dataset,nAP50 increases by 4.0 to 29.4.