摘要
针对现有的弱监督目标检测算法由于缺乏实例级类别的注释,易出现局部定位的问题,提出一种基于空间-通道注意力机制与多实例优化回归网络相结合的弱监督目标检测算法。通过在特征提取网络中引入注意力模块,发掘出更为优质的初始伪真值标签,有效地提取了隐含的位置信息。在网络训练阶段引入自适应的策略挖掘出训练细化分支的有效监督,实现对卷积神经网络中实例分类器的优化,同时以端到端的方式进行模型的训练,避免网络过多地关注目标的显著区域而不是整个对象,从而使模型跳出局部最优,提升模型的检测性能。在PASCAL VOC 2007和VOC 2012大规模数据集上的实验结果表明,提出的算法拥有比近几年主流方法更好的检测性能,有效缓解了局部定位的问题。
Aiming at the problem that existing weakly supervised object detection algorithms are prone to localization due to the lack of instance-level annotations,we proposed a weakly supervised object detection algorithm based on the combination of space-channel attention mechanism and multi-instance optimized regression network.By introducing the attention module in the feature extraction network,better initial pseudo-truth labels are discovered,and the hidden location information is effectively extracted.Introduce an adaptive strategy to mine the effective supervision of the training refinement branch,realize the optimization of the instance classifier in the convolutional neural network,and jointly train the multi-instance detection branch and the regression branch sharing the same backbone in an end-to-end manner,avoiding network pay too much attention to the salient area of the target instead of the entire object,so that the model can jump out of the local optimum and improve the detection performance of the model.The experimental results on the PASCAL VOC 2007 and VOC 2012 large-scale data sets show that the algorithm proposed in this paper has better detection performance than the mainstream methods in recent years,and effectively alleviates the problem of local positioning.
作者
杨振文
葛斌
郑海君
邬成
YANG Zhenwen;GE Bin;ZHENG Haijun;WU Cheng(College of Computer Science and Engineering,Anhui University of Science and Technology,Huainan Anhui 232001,China)
出处
《佳木斯大学学报（自然科学版）》
CAS
2023年第6期12-17,共6页
Journal of Jiamusi University：Natural Science Edition
基金
国家自然科学基金(6210071479,62102003)
国家重大专项(2020YFB1314103)
安徽省自然科学基金(2108085QF258)
安徽省博士后基金(2022B623)
安徽省高等学校自然科学研究项目(KJ2020A0299)。
关键词
目标检测
深度学习
弱监督学习
注意力机制
自适应学习
object detection
deep learning
weakly supervised learning
attention mechanism
adaptive learning
