面向连续图像序列的语义分割  

Semantic segmentation for continuous image sequence

作　　者：邹序焱 何汉武[1,2] 吴悦明[1] ZOU Xuyan;HE Hanwu;WU Yueming(School of Electromechanical Engineering,Guangdong University of Technology,Guangzhou 510006,China;Guangdong Polytechnic of Industry and Commerce,Guangzhou 510510,China;Department of Artificial Intelligence and Big Data,Yibin University,Yibin 644007,China)

机构地区：[1]广东工业大学机电工程学院,广东广州510006 [2]广东工贸职业技术学院,广东广州510510 [3]宜宾学院人工智能与大数据学部,四川宜宾644007

出　　处：《现代电子技术》2022年第14期127-131,共5页Modern Electronics Technique

基　　金：国家重点研发专项(2018YFB1004902)。

摘　　要：为弥补传统语义分割模型忽略时间维度上信息的不足,文中提出一种时间连续的语义分割模型。该模型在经典的U-Net语义分割模型基础上,利用前几帧分割的结果,添加时间维度上的图像特征模块;再通过特征融合模块对各部分的特征图进行融合;最后,利用融合的特征对像素进行分类。为验证所提出模型的分割效果,采集连续图像数据集并利用Labelme对其进行标注,在该数据集上进行训练和测试。结果表明,在同等条件下,基于时间连续的语义分割模型在Dice系数、精确率和召回率方面均高于U-Net网络,对运动模糊图像的语义分割效果也较好,说明该方法能够改善语义分割效果。In order to make up for the deficiency that the traditional semantic segmentation model ignores the information in the time dimension,a time continuous semantic segmentation model is proposed in this paper. Based on the classical U-Net semantic segmentation model,the segmentation results of the previous frames are used to add the image feature module in the time dimension,and then feature fusion modules are used to fuse the feature images of each part. The fused features are used to classify pixels. In order to verify the semantic effect of the segmentation model,the continuous image dataset is collected and labeled with Labelme,and the dataset training and testing are conducted. The results show that under the same conditions,the Dice coefficient,accuracy and recall rate of the semantic segmentation model based on time continue are higher than that of UNet network,and the semantic segmentation effect of motion blurred image is also better than U-Net mode,which shows that the method can improve the semantic segmentation effect.

关 键 词：语义分割 图像序列 深度学习 相似性 图像处理 卷积神经网络 损失函数 数据标注 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP751[电子电信—信息与通信工程]