《计算机应用研究》|Application Research of Computers

使用九点双三次卷积插值方法改进的DeepLab-v3模型

Improved DeepLab-v3 model using nine-point bicubic convolution interpolation

免费全文下载 (已被下载 次)  
获取PDF全文
作者 宋国杰,黄佳芳,陈普春,陈亚丽
机构 西南石油大学 理学院,成都 610500
统计 摘要被查看 次,已被下载
文章编号 1001-3695(2020)09-065-2876-05
DOI 10.19734/j.issn.1001-3695.2019.04.0125
摘要 针对DeepLab-v3模型中双线性插值方法只能从邻近的四个抽样点提取信息,导致图像分割的精确度下降问题,提出两个改进的DeepLab-v3模型,即IDL-v3-1(improved DeepLab-v3-1 model)和IDL-v3-2(improved DeepLab-v3-2 model)。两个改进的模型使用九点双三次卷积插值方法替换DeepLab-v3模型的双线性插值方法,以获得更精确的分割图像。在PASCAL VOC 2012数据集上检验发现,与DeepLab-v3模型相比,IDL-v3-1和IDL-v3-2的精确度分别提升了0.43%和0.59%,分割每张图像的耗时量分别增加了0.002 s和0.005 s。实验结果表明,改进的DeepLab-v3模型可以在与DeepLab-v3模型相近的时间成本内达到更高的图像分割精度。
关键词 图像语义分割; DeepLab-v3; 双线性插值; 双三次卷积插值
基金项目 国家自然科学基金面上项目(41674141)
本文URL http://www.arocmag.com/article/01-2020-09-065.html
英文标题 Improved DeepLab-v3 model using nine-point bicubic convolution interpolation
作者英文名 Song Guojie, Huang Jiafang, Chen Puchun, Chen Yali
机构英文名 School of Science,Southwest Petroleum University,Chengdu 610500,China
英文摘要 In view of the fact that the bilinear interpolation method in the DeepLab-v3 model can only extract information from four adjacent sampling points, resulting in the degradation of image segmentation accuracy, this paper proposed two improved DeepLab-v3 models: IDL-v3-1 and IDL-v3-2. The two improved models replaced the bilinear interpolation method of the DeepLab-v3 model with a nine-point bicubic convolution interpolation method to get a more accurate segmentation image. It tested model on the PASCAL VOC 2012 dataset, it was found that compared with the DeepLab-v3 model, the accuracy of IDL-v3-1 and IDL-v3-2 were increased by 0.43% and 0.59%, respectively, and the time taken to segment each image was increased by 0.002 s and 0.005 s, respectively. Experimental results show that the improved DeepLab-v3 model can achieve higher image segmentation accuracy in a similar time cost as the DeepLab-v3 model.
英文关键词 image semantic segmentation; DeepLab-v3; bilinear interpolation; bicubic convolution interpolation
参考文献 查看稿件参考文献
 
收稿日期 2019/4/24
修回日期 2019/6/18
页码 2876-2880
中图分类号 TP183
文献标志码 A