基于孪生网络视频烟雾目标分割方法  

Video smoke target segmentation method based on twin network

作　　者：刘方涛 杨剑[1] 白瑞君 张启尧 任宇杰 LIU Fang-tao;YANG Jian;BAI Rui-jun;ZHANG Qi-yao;REN Yu-jie(Software School,North University of China,Taiyuan 030051,China)

机构地区：[1]中北大学软件学院,山西太原030051

出　　处：《计算机工程与设计》2020年第9期2597-2603,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61602427);山西省回国留学人员科研基金项目(2014-053)。

摘　　要：烟雾作为不规则目标且识别窗口长宽比不固定,目标跟踪框中有一定误差。在目标分割领域中烟雾的流动性可以被很好地标识出来,且为解决在线训练的费时问题和提高识别速度,对完全卷积Siamese方法进行研究,提出基于ResNet50的孪生网络对烟雾视频数据离线训练方法,输出回归预测score(分数)和mask(掩膜),由于mask不好预测,用一个ROW来预测mask,达到目标分割情况下的跟踪。Smoke as an irregular target,the recognition window aspect ratio is not fixed,there is an error in the target tracking frame.The liquidity of smoke in the target segmentation field can be well identified.To solve the time-consuming problem of online training and improve the recognition speed,the fully convoluted Siamese method was researched,and the offline training method of smoke video data based on ResNet50 was proposed.The regression prediction score and mask were outputted.Since the mask was not well predicted,a ROW was used,and mask was prediceted.The tracking in the case of target segmentation was achieved.

关 键 词：半监督 视频对象分割 全卷积 二元分段任务 双输入 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]