摘要
针对当前Deeplab v3+模型没有充分采用高分辨率的浅层特征出现的错误分割、遗漏分割等现象,提出一种融合多尺度特征的改进Deeplab v3+特征图像语义分割算法。在主干网络中,引入多尺度金字塔卷积;将空洞空间卷积池化金字塔中的标准卷积替换为深度可分离卷积,减少整体模型的参数量;最后,在解码层采用多尺度方法来捕捉获取全局背景,将背景特征通过注意力机制,再与浅层特征和空洞空间金字塔池化层结合,丰富融合后的浅层特征语义信息。实验表明,在CityScapes验证集中,所提算法具有更好的边缘分割效果,平均交并比达到了74.76%,较原有算法提升了2.20%。通过与先进算法比较,也证明所提算法应对改善错误分割、遗漏分割的有效性。
To address the phenomena of incorrect segmentation and missing segmentation that occur when the current Deeplab v3+model does not adequately employ high-resolution shallow features an improved Deeplab v3+feature image semantic segmentation algorithm that incorporates multi-scale features is proposed.In the backbone network multi-scale pyramidal convolution is introduced.The standard convolution in the pooled pyramid of atrous space convolution is replaced by the deep separable convolution to reduce the number of parameters of the whole model.Finally a multi-scale approach is adopted in the decoding layer to capture the global background and the background features are combined with the shallow features and the atrous space pyramid pooling layer through the attention mechanism to enrich the semantic information of the fused shallow features.Experiments show that in CityScapes dataset the proposed algorithm has a better edge segmentation effect with an Mean Intersection over Union(MIoU)of 74.76%which is 2.20%higher than that of the original algorithm.Compared with advanced algorithms it is also proved that it is effective in improving incorrect segmentation and missing segmentation.
作者
张文博
瞿珏
王崴
胡俊
王庆力
ZHANG Wenbo;QU Jue;WANG Wei;HU Jun;WANG Qingli(Air Force Engineering University,Xi'an 710000 China)
出处
《电光与控制》
CSCD
北大核心
2022年第11期12-16,30,共6页
Electronics Optics & Control
基金
国家自然科学基金(52175282)。
关键词
深度学习
语义分割
多尺度
注意力机制
迁移学习
deep learning
semantic segmentation
multi-scale
attention mechanism
transfer learning
