特征自蒸馏机制下的弱监督目标检测

Weakly Supervised Object Detection Based on Feature Self-Distillation Mechanism

目前基于图像级注释信息的主流弱监督目标检测算法常常出现局部定位问题,仅仅关注图像中局部高辨别性的区域,却忽略了完整的目标。为了解决这种问题,提出了一种端对端的基于特征自蒸馏的弱监督目标检测网络(FSDNet),其中可拆卸的特征自蒸馏模块充分利用不同层级特征表示中的细节信息和语义信息,并通过特征自蒸馏损失约束网络训练,在未增加测试期计算代价的前提下增强了检测器综合性能;同时构造回归分支简单却有效地提取并利用特征中隐性位置信息,并通过改善监督信息生成算法、平衡优化损失等策略进一步改善了弱监督目标检测器的局部定位问题。在Pascal VOC 2007、VOC 2012、MS-COCO等大规模公开数据集上的实验结果表明,FSD-Net拥有比Baseline及近年主流方法更好的检测性能,有效地缓解了局部定位问题。

The current mainstream weakly supervised object detection methods based on imagelevel annotation often occur local localization problem,tend to overfit the most discriminative regions,and ignore object integrity.To solve these existing problems,an endtoend weakly supervised object detection network based on feature selfdistillation(FSDNet),in which the detachable feature selfdistillation module fully uses the semantic and detailed information in the representation of different hierarchical features,is proposed.Additionally,through feature selfdistillation loss constraint network training,the comprehensive performance of the detector is enhanced without increasing the calculation cost during the test period.Moreover,the regression branches are constructed to simply extract and effectively utilize the implicit location information in the features,improves the original supervision information generation algorithm,and balances optimization loss and other strategies to further improve the local localization problem of the weakly supervised object detector.Experiments on largescale public datasets,such as Pascal VOC 2007,VOC 2012,and MSCOCO,show that FSDNet has a better detection performance than the Baseline and other existing mainstream methods,effectively alleviating the local localization problem in weakly supervised object detection.