吴稳稳,吴晓红,刘强,等. 基于全局注意力的多级特征融合目标检测算法[J]. 科学技术与工程, 2020, 20(27): 11185-11191.
WU Wen-wen,WU Xiao-hong,Liu Qiang,et al.Multi-level Feature Fusion Object Detection Algorithm Based on Global Attention[J].Science Technology and Engineering,2020,20(27):11185-11191. |
| 基于全局注意力的多级特征融合目标检测算法 |
| Multi-level Feature Fusion Object Detection Algorithm Based on Global Attention |
| 投稿时间：2019-12-13 修订日期：2020-09-14 |
| DOI： |
| 中文关键词: 卷积神经网络 目标检测 注意力增强 多级特征融合 |
| 英文关键词:convolutional neural networks object detection attention enhancement multi-level feature fusion |
| 基金项目:国家自然科学基金项目（面上项目，重点项目，重大项目）,四川省科技厅国际科技合作与交流研发项目（2018HH0143）、成都市产业集群协同创新项目（2016-XT00-00015-GX） |
|
| 摘要点击次数: 172 |
| 全文下载次数: 76 |
| 中文摘要: |
| 针对小目标物体检测精度差的问题,同时不以牺牲速度为代价,本文提出了一种基于全局注意力的多级特征融合目标检测算法。算法首先由卷积神经网络生成多尺度的特征图,然后采用多级特征融合的方法,将浅层和深层特征图的语义信息相结合,提高特征图的表达能力,接着引入全局注意力模块,对特征图上下文信息进行建模,并捕获通道之间的依赖关系来选择性地增强重要的通道特征。此外,在多任务损失函数的基础上增加一项额外的惩罚项来平衡正负样本。最后经过分类回归、迭代训练和过滤重复边框得到最终检测模型。对所提算法在PASCAL VOC数据集上进行了训练和测试,结果表明该算法能有效地提升小目标物体检测效果,并较好地平衡了检测精度与速度之间的关系。 |
| 英文摘要: |
| Aiming at the problem of poor detection accuracy of small target objects without sacrificing speed, this paper proposes a multi-level feature fusion object detection algorithm based on global attention. The algorithm first generated a multi-scale feature map from a convolutional neural network, and then used a multi-level feature fusion method to combine the semantic information of the shallow and deep feature maps to improve the expression ability of the feature map. Then, the global attention module was introduced to model the feature map context information and captured the dependencies between channels to selectively enhance important channel features. In addition, an additional penalty term was added to the multi-task loss function to balance the positive and negative samples. Finally, the final detection model was obtained through classification regression, iterative training, and filtering repeated borders. The proposed algorithm was trained and tested on the PASCAL VOC dataset, and the results show that the algorithm can effectively improve the detection effect of small targets and balance the relationship between detection accuracy and speed. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
|
|
