摘要
传统的目标检测方法不能有效检测微操作系统中部分受遮挡或多种姿态的目标,因此文中采用改进的基于区域卷积神经网络的Faster-RCNN检测算法,用于微操作系统中部分受遮挡或多种姿态的目标检测.在原始Faster-RCNN的基础上,使用在图像分类任务中性能优越的深度残差网络作为检测算法的主框架,并且引入防止正负样本不均衡的在线困难样本挖掘策略以提高网络性能.实验表明,这种改进的基于区域卷积神经网络方法能有效识别部分受遮挡和不同姿态的目标,相比传统方法,文中方法对环境适应性更强,速度更快,具有实际应用价值.
In micro-operating system, traditional object detection method cannot detect the objects with partial occlusion and multiple poses, and thus an improved faster region convolutional neural network （Faster RCNN） is adopted to solve the problem. On the basis of original Faster RCNN, deep residual network exhibiting excellent performance in image classification is introduced as the framework of the algorithm, and online hard example mining strategy to enhance the performance by alleviating the imbalance between positive and negative examples is employed. The experimental results manifest that the proposed method can detect objects with partial occlusion and multiple poses effectively. The proposed method shows strong adaptability to environment, responds quickly compared with traditional methods, and thus the practicality of it is verified.
出处
《模式识别与人工智能》
EI
CSCD
北大核心
2018年第2期142-149,共8页
Pattern Recognition and Artificial Intelligence
基金
国家高科技研究发展计划(863计划)(No.2008AA8041302)
国家自然科学基金项目(No.60275013)资助~~
关键词
微操作系统
目标检测
区域卷积神经网络
深度残差网络
在线困难样本挖掘
Micro-operating System, Object Detection, Region Convolutional Neural Network, DeepResidual Network, Online Hard Example Mining
