多重边界参考的弱监督语义分割网络
Weakly Supervised Semantic Segmentation Network with Multiple Boundary References
-
摘要: 针对弱监督语义分割任务中目标被重叠或遮挡,种子区域生成的伪像素掩模难以准确覆盖目标区域,出现的漏分割和错分割问题,提出一种多重边界参考的弱监督语义分割网络.首先设计了一个边界探索模块,该模块聚焦于融合空间域边界和频率域边界,并以空间域和频率域互参考方式得到目标区域边界,使网络可以在目标重叠或遮挡的情况下仍可以准确找到目标;然后,将得到的目标区域边界作为参考信息输入网络中,使网络可以生成覆盖更准确目标区域的伪像素掩模,进而提高网络在目标重叠或遮挡情况下的分割精度.在通用数据集PASCAL VOC2012和COCO 2014上对其进行实验验证和分析,所提网络的mIoU值分别达到了69.2%和40.6%,并可以有效地改善目标被遮挡或重叠时分割精度低的问题,提升弱监督语义分割精度.Abstract: In response to the challenges of object occlusion and overlap in weakly supervised semantic segmentation tasks, we propose a multi-boundary-reference weakly supervised semantic segmentation network. The network incorporates a boundary exploration module, which focuses on integrating spatial-domain and frequency-domain boundaries and obtains object region boundaries through mutual referencing between spatial and frequency domains. This enables the network to accurately locate objects even in cases of overlap or occlusion. Subsequently, the obtained object region boundaries are inputted into the network as reference information, allowing the network to generate pseudo pixel masks that more accurately cover the target region. This, in turn, enhances the network’s segmentation accuracy in scenarios of object overlap or occlusion. Experimental validation and analysis conducted on the widely used datasets PASCAL VOC 2012 and COCO 2014 demonstrate that the proposed network achieves mIoU values of 69.2% and 40.6%, respectively. Furthermore, it effectively addresses the issue of low segmentation accuracy in cases of object occlusion or overlap, thereby improving the accuracy of weakly supervised semantic segmentation.