|
| MSA-Net:一种基于多阶段注意力机制的少样本目标检测方法 |
| MSA-Net: few-shot object detection with multi-stage attention mechanism |
| 投稿时间:2023-02-03 |
| DOI:10.3969/j.issn.1005-5630.202302030011 |
| 中文关键词: 深度学习 少样本学习 目标检测 |
| 英文关键词:deep learning few-shot learning object detection |
| 基金项目:基础科研条件与重大科学仪器设备研发计划(2022YFF0706003) |
|
| 摘要点击次数: 2368 |
| 全文下载次数: 1446 |
| 中文摘要: |
| 近年来,样本较少场景下的目标检测引起了广泛的关注。由于少样本提供的信息有限,大部分少样本目标检测模型采用改进的Faster RCNN检测框架进行研究。但由于Faster RCNN框架中潜在的模块矛盾问题,现有的少样本目标检测模型的特征捕捉和分类的能力有待提高。为解决以上问题,以Faster RCNN框架为基础,加入了梯度反传解耦机制,缓解在反向传播过程中,RPN和RCNN的冲突对主干网络的负面影响。为提高目标检测模型的特征捕捉能力,采用元学习框架,并融合基于注意力机制的蒸馏模块和多尺度注意力模块,充分利用查询集和支持集的信息,捕捉更多全局特征信息。大量的实验证明,在随机采样目标数k=1, 2, 3, 5, 10设置下,改进后的模型在Pascal VOC数据集的新类上,分别达到21.8%,34.7%,40.9%,44.5%,51.7% mAP(AP50)。在k = 10, 30设置下,改进后的模型在COCO数据集的新类上,分别达到25.1%,27.6% mAP(AP50)。 |
| 英文摘要: |
| In recent years, object detection in scenarios with fewer samples has attracted widespread attention. Due to the limited information provided by the few samples, most of few-shot object detection models are studied using the improved Faster RCNN detection framework. However, due to the potential module contradiction problem in the Faster RCNN framework, the feature capture and classification capabilities of the existing few-shot object detection models need to be improved. In order to solve the above problems, this paper adds a gradient decoupling mechanism based on the Faster RCNN framework to alleviate the negative impact of the conflict between RPN and RCNN on the backbone network during the backpropagation process. In order to improve the feature detection ability of the object detection model, this paper adopts meta-learning framework, integrates the distillation module based on attention mechanism and the multi-scale attention module, and makes full use of the information of the query set and support set to capture more global feature information. A large number of experiments have proved that under the setting of randomly sampled shot amount k=1, 2, 3, 5, 10, the improved model can reach 21.8%, 34.7%, 40.9%, 44.5%, 51.7% mAP (AP50) on the new class of Pascal VOC dataset, respectively. Under the k=10, 30 setting, the improved model achieves 25.1% and 27.6% mAP (AP50) on the new class of the COCO dataset, respectively. |
| HTML 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|