Page 338 - 《软件学报》2020年第11期

李阳 等:基于对象位置线索的弱监督图像语义分割方法 3653

类别“plant”占据图像的区域较小,常常被放置在角落,并且轮廓复杂,因此也同样不容易被分割出来.
Table 7 Our segmentation results for each class on validation and test sets
表 7 本文方法在验证和测试集上针对每个类别的分割结果
WSS_S 2_AM_DeepLab DHSN_S 2_AM_DeepLab WSS_S 2_AM_DeepLab DHSN_S 2_AM_DeepLab
类别
验证集(%) 验证集(%) 测试集(%) 测试集(%)
background 85.0 86.8 85.8 87.2
airplane 69.0 73.7 68.7 74.3
bike 25.5 26.3 28.9 30.9
bird 67.9 67.6 67.6 71.2
boat 49.6 57.2 39.3 44.8
bottle 62.1 65.9 57.5 61.6
bus 72.9 73.6 70.5 72.2
car 61.3 66.6 61.4 64.1
cat 70.6 70.9 66.5 70.5
chair 18.2 13.5 21.2 16.7
cow 57.4 63.4 57.2 58.2
diningtable 32.1 14.9 34.5 22.8
dog 60.4 63.2 65.0 67.1
horse 57.4 59.3 59.7 59.2
motorbike 60.9 63.3 67.4 68.6
person 45.1 59.2 47.3 60.0
plant 30.5 34.0 39.5 37.4
sheep 65.7 65.0 65.9 68.7
sofa 29.0 20.5 34.7 22.1
train 59.4 59.3 52.3 57.1
tv/monitor 39.8 48.1 41.9 46.2
Average 53.3 54.9 53.9 55.3
图 8 展示了本文所提方法的分割结果.我们可以看出,方法“DHSN_S 2 _AM_DeepLab”的分割结果更加准确.

Fig.8 Visual segmentation results of the proposed methods on PASCAL VOC 2012 validation set
图 8 本文方法在 PASCAL VOC 2012 验证集上的可视化分割结果