结合注意力机制的深度学习光流网络  

Deep Optical Flow Learning Networks Combined with Attention Mechanism

作　　者：周海赟[1] 项学智[2] 翟明亮 张荣芳 王帅 ZHOU Haiyun;XIANG Xuezhi;ZHAI Mingliang;ZHANG Rongfang;WANG Shuai(Institute of Public Security,Nanjing Forest Police College,Nanjing 210023,China;School of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]南京森林警察学院治安学院,南京210023 [2]哈尔滨工程大学信息与通信工程学院,哈尔滨150001

出　　处：《计算机科学与探索》2020年第11期1920-1929,共10页Journal of Frontiers of Computer Science and Technology

基　　金：中央高校基本科研业务费专项资金No.LGZD201902;国家自然科学基金No.61401113;黑龙江省自然科学基金No.LC201426。

摘　　要：为提升基于编解码架构的U型网络在深度学习光流估计中的精度,提出了一种结合注意力机制的改进有监督深度学习光流网络。网络由收缩和扩张两部分组成,收缩部分利用一系列卷积层来提取图像之间的高级特征,扩张部分通过反卷积操作将特征图恢复至原始图像分辨率,将通道注意力机制引入U型网络架构中以学习通道之间的相互依赖性,自适应地调整各通道的特征权重,增强网络的特征提取能力。同时,改进的网络还使用了空洞卷积以在卷积核尺寸不变的情况下增大感受野,使用变分光流方法中的恒常约束与平滑约束以进一步利用运动先验知识提升估计效果。最后基于合成图像序列数据集进行了实验验证,实验结果表明所设计的网络能够有效提升深度学习光流估计的准确率。In order to improve the accuracy of deep learning optical flow estimation based on encoder-decoder UNet,a modified supervised deep optical flow learning network combined with attention mechanism is proposed,which consists of a contracting part and an expanding part.In contracting part,high-level feature information is extracted using a series of convolutional layers,and spatial feature maps are then restored to full resolution by conducting successive deconvolution in expanding part.In this paper,attention mechanism is embedded in U-Net to learn interdependencies among the channels so that the channel-wise features can be weighted adaptively,which can enhance the performance of feature extraction.Meanwhile,the proposed network also combines dilated convolution to enlarge the receptive field without changing the size of convolutional kernel.Further,constancy constraints and smoothness constraints from variational method are also adopted so that priori knowledge can be used to improve the accuracy of optical flow estimation.Extensive experiments are conducted on synthesis image sequence datasets and the experimental results show the proposed network is effective for improving accuracy of deep learning optical flow estimation.

关 键 词：光流估计 深度学习 注意力机制 空洞卷积 先验约束 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]