基于特征融合的端到端弱监督语义分割

End-to-end Weakly Supervised Semantic Segmentation Based on Feature Fusion

弱监督语义分割是一项极其具有挑战性的任务,它仅以图像级标签作为监督。现有的方法大都将骨干模型最后一层卷积层特征图与分类器权重相乘得到最具有判别力的区域,然而,仅仅得到最显著的特征对语义分割任务来说是不够的。遵循融合多层特征具有更加丰富的语义信息这一事实,将不同层的特征图进行融合,对融合后的特征图进行目标对象定位。将该方法迁移到一个端到端的网络框架中,语义分割性能显著提升。该端到端网络在Pascal VOC2012数据集的验证集上达到了62.5%的结果。

Weakly supervised semantic segmentation is an extremely challenging task that is supervised only with image-level labels.Most of the existing methods multiply the last convolutional layer feature map of the backbone model with the classifier weights to get the most discriminative region.However,just getting the most salient features is not enough for the semantic segmentation task.In this paper,following the fact that fused multi-layer features have richer semantic information,the feature maps of different layers are fused and the fused feature maps are used for target object localisation.Migrating this approach to an end-to-end network framework,the semantic segmentation performance is significantly improved.The end-to-end network in this paper achieves 62.5%results on the validation set of Pascal VOC2012 dataset.