分解多空洞深度卷积的轻量级图像语义分割  

Lightweight Image Semantic Segmentation Model Based on Factorizing Multi-hole Deep Convolution

作　　者：宣明慧 张荣国[1] 李富萍[1] 赵建[1] 胡静[1] XUAN Ming-hui;ZHANG Rong-guo;LI Fu-ping;ZHAO Jian;HU Jing(School of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China)

机构地区：[1]太原科技大学计算机科学与技术学院,太原030024

出　　处：《太原科技大学学报》2022年第3期191-196,共6页Journal of Taiyuan University of Science and Technology

基　　金：国家自然科学基金(51375132);山西省自然科学基金(201801D121134)。

摘　　要：为了降低图像语义分割网络的模型复杂度,提出了分解多空洞深度卷积的轻量级图像语义分割模型。首先,针对大小不一的多目标,用感受野不同的金字塔结构提取图像语义特征,在深度卷积过程中对空洞卷积进行分解,以降低参数量和计算量;其次,对不同阶段得到的特征图进行融合,利用子像素卷积进行上采样,将提取的低分率图插入到最后输出的高分辨率特征图中,以改善图像语义分割的精准性;最后,在CamVid数据集上和现有的9种方法进行对比实验,结果表明,该方法参数量和计算量都相对较低,同时语义分割精准性得以提升。In order to reduce the model complexity of image semantic segmentation network,this article proposes a lightweight image semantic segmentation model that decomposes multi-hole deep convolution.First,for multiple targets of different sizes,the image semantic features are extracted with different pyramid structures in the receptive field,and the Atrous Convolution is decomposed in the Depthwise process to reduce the amount of parameters and calculations,and the real-time inference is realized;Secondly,the feature maps obtained at different stages are fused,and sub-pixel convolution is used for upsampling,and the extracted low-rate maps are inserted into the final output high-resolution feature maps to improve the accuracy of image semantic segmentation;Finally,comparing experiments with nine existing methods on the CamVid dataset,the results show that the parameters and calculations of this method are relatively low,and the accuracy of semantic segmentation is improved.

关 键 词：图像语义分割 金字塔结构 空洞卷积 深度可分离卷积 分解卷积 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]