基于双通道卷积神经网络的深度图超分辨研究  

Depth Map Super-Resolution Based on Two-Channel Convolutional Neural Network

作　　者：李素梅 雷国庆 范如 Li Sumei;Lei Guoqing;Fan Ru(School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China)

机构地区：[1]天津大学电气自动化与信息工程学院,天津300072

出　　处：《光学学报》2018年第10期128-134,共7页Acta Optica Sinica

基　　金：国家自然科学基金(61002028);国家863计划(2012AA011505;2012AA03A301)

摘　　要：目前直接获取的深度图受其成像原理及硬件设备等因素的限制,存在分辨率低、边缘信息丢失等缺点,极大地影响了深度图的应用。针对这一问题,提出基于双通道卷积神经网络的深度图超分辨率重建模型。该模型由深、浅两个通道组成,21层的深层通道通过联合卷积与反卷积,结合跳跃连接与多尺度理论,实现深度图细节特征的快速学习;3层的浅层通道用于学习深度图的轮廓特征;最后融合深、浅两个通道,将细节与轮廓相结合,实现由低分辨率深度图到高分辨率深度图的端到端的学习。该模型充分利用卷积神经网络的学习能力自主提取深度图的有效特征,避免了手工提取特征的不准确性。在Middlebury RGBD数据集上的实验结果表明,本文模型在大采样因子8时仍能取得较好的效果,具有很高的实际应用价值。The depth map obtained directly is limited by the disadvantages such as low resolution and missing edge information,it greatly affects the application of depth map.In order to solve this problem,a two-channel convolutional neural network for depth map super-resolution is proposed.It consists of two channels,deep and shallow,and there are 21 layers in the deep network.Through joint convolution and deconvolution,combining skip connection and multi-scale theory,the deep channels can quickly learn the detailed features of depth map.Shallow network of 3 layers are used to learn the rough features of depth maps.Finally,the two channels are combined with details and outlines to realize end-to-end mapping from low resolution depth map to high resolution one.The model makes full use of the learning ability of the convolutional neural network to independently extract the effective features of the depth map and avoid the inaccuracy of manually extracting features.The experimental results on the Middlebury RGBD dataset show that the proposed model can achieve good results at a large sampling factor of 8,and has a high practical value.

关 键 词：图像处理 超分辨率 深度图 卷积神经网络 残差网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]