%0 Journal Article
%T 基于LAE-DeepLabv3+的舌象图像分割方法
Tongue Image Segmentation Method Based on LAE-DeepLabv3+
%A 李月月
%A 石文
%A 曾庆红
%A 张静宇
%A 魏楚婷
%A 王皓宇
%A 孙雅星
%A 郑涛
%J Software Engineering and Applications
%P 190-204
%@ 2325-2278
%D 2026
%I Hans Publishing
%R 10.12677/sea.2026.152019
%X 针对舌象图像分割中模型参数量大、舌体边缘模糊及易受背景干扰等问题，提出一种轻量化高精度分割模型LAE-DeepLabv3+。首先，采用EfficientViT替换主干网络，大幅降低模型计算复杂度；其次，在低层特征分支引入EMA (Efficient Multi-scale Attention)模块，增强网络对舌体边界及局部细节的感知能力；最后，在ASPP (Atrous Spatial Pyramid Pooling)模块中增加水平与垂直条带池化分支，强化方向性上下文与长程依赖建模。在自建舌象数据集上的实验结果表明，该模型的IoU和Dice分别达到96.47%和98.20%，参数量仅为12.12 M。该方法有效实现了分割精度与轻量化的平衡，为舌象客观化分析提供了可靠支持。
To address the challenges of large parameter volumes, blurred tongue edges, and background interference in tongue image segmentation, a lightweight and high-precision segmentation model, LAE-DeepLabv3+, is proposed. First, the EfficientViT architecture is employed to replace the original backbone network, significantly reducing computational complexity. Second, an EMA (Efficient Multi-scale Attention) module is integrated into the low-level feature branch to enhance the network's perception of tongue boundaries and local details. Finally, horizontal and vertical strip pooling branches are added to the ASPP (Atrous Spatial Pyramid Pooling) module to strengthen directional context and long-range dependency modeling. Experimental results on a self-built tongue image dataset demonstrate that the model achieves an IoU of 96.47% and a Dice coefficient of 98.20%, with a parameter count of only 12.12 M. This method effectively balances segmentation accuracy and model lightness, providing reliable support for the objective analysis of tongue diagnosis.
%K 舌象分割，
%K DeepLabv3+，
%K 轻量化网络，
%K EMA模块，
%K 多尺度特征融合
Tongue Image Segmentation
%K DeepLabv3+
%K Lightweight Network
%K EMA Module
%K Multi-Scale Feature Fusion
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=140133