摘要
由于缺少精确的边界框注释,弱监督目标检测器依赖预训练图像分类模型对候选区域进行分类.然而,预训练模型通常对具有鉴别性的区域而非完整的目标产生高响应,导致局部主导、实例丢失和非紧密框等问题.为此,文中提出基于多层次融合的弱监督目标检测网络,从增强对弱鉴别性空间特征的学习、类内样本特征丰富性和可信伪标签权重的角度提升检测性能.首先,幂池化层利用幂函数加权融合邻域内的激活值,减少弱鉴别性特征的信息损失.其次,特征混合方法随机融合候选区域的特征向量,丰富训练样本特征的多样性.最后,基于置信度的样本重加权策略融合预测值和伪标签的置信度,调节伪标签对训练的影响.在3个基准数据集上的实验表明文中网络性能较优.
Due to the lack of precise bounding box annotations,weakly supervised object detectors rely on the pretrained image classification model to classify candidate regions.However,the pretrained model often produces high responses for discriminative regions rather than complete objects,resulting in the problems of part domination,instance missing and untight boxes.To address these issues,a multi-level fusion based weakly supervised object detection network is proposed.The detection performance is improved from the perspectives of enhancing the weak discriminative spatial feature learning,enriching intra-class sample features and weighting reliable pseudo-labels.Firstly,a power function is utilized to weight and fuse the activation values within the neighborhood by the power pooling layer to reduce information loss of weak discriminative features.Secondly,the feature vectors of candidate regions are randomly fused by the feature mixing method to enrich the diversity of training sample features.Finally,the confidence of predictions and pseudo-labels is fused via the confidence-based sample re-weighting strategy to adjust the influence of pseudo-labels on training.Experiments on three benchmarks demonstrate the superiority of the proposed network.
作者
曹环
陈曾平
CAO Huan;CHEN Zengping(School of Electronics and Communication Engineering,Sun Yat-Sen University,Shenzhen 518107)
出处
《模式识别与人工智能》
EI
CSCD
北大核心
2024年第5期424-434,共11页
Pattern Recognition and Artificial Intelligence
关键词
目标检测
弱监督学习
多层次融合
深度网络
Object Detection
Weakly Supervised Learning
Multi-level Fusion
Deep Network
