基于深度可分离卷积的级联多尺度交叉网络的超分辨率重构  

Super-Resolution Reconstruction of Cascaded Multi-Scale Crossover Networks Based on Depthwise Separable Convolutions

作　　者：商丽娟 应自炉[1] 徐颖[1] 

机构地区：[1]五邑大学信息工程学院,广东江门

出　　处：《图像与信号处理》2018年第2期96-104,共9页Journal of Image and Signal Processing

基　　金：国家基金项目No.61372193,No.61771347;广东高校优秀青年教师培训计划资助项目No.SYQ2014001;广东省特色创新项目No.2015KTSCX143,2015KTSCX145,2015KTSCX148;广东省青年创新项目No.2015KQNCX172,No.2016KQNCX171;江门市科技计划项目No.201501003001556,No.201601003002191.

摘　　要：深度卷积神经网络进行单幅图像超分辨率在精度和速度方面取得了显著成果。然而随着网络的层次加深,信息流动减弱,训练难以实现。同时,大多数模型采用单流结构,在不同的感受域下的上下文信息很难获取。改善信息流动获取足够的细节信息并减少网络参数量,本文提出了基于深度可分离卷积的级联多尺度交叉网络(Cascaded Multiscale Crossing Network, CMSC),在每个级联子网络中,堆叠多个多尺度交叉模块以便融合互补多尺度的信息,从而有效改善跨层的信息流。同时,在每个阶段引入残差学习策略,充分利用低分辨率特征信息,进一步提升重建性能。在基准数据集的评估表明本文方法优于最主流的超分辨率方法。Deep convolutional neural networks have achieved remarkable results in accuracy and speed for single image super-resolution. However, when the depth of the network is increased, the flow of information is weakened and training is difficult to achieve. At the same time, most existing net-work structures use a single-stream structure, and context information is difficult to obtain under different receptive domains. To improve information flow to get enough details and reduce network parameters, in this paper, a Cascaded Multiscale Crossing Network (CMSC) based on a deeply separable network is proposed. In each cascading sub-network, multiple multi-scale crossover modules are stacked to fuse complementary multi-scale information, thereby effectively improving cross-layer information flow. At the same time, a residual learning strategy was introduced at each stage to make full use of the low-resolution feature information to further enhance the re-construction performance. The evaluation of the benchmark datasets shows that the proposed method outperforms the most mainstream superresolution methods.

关 键 词：单图像超分辨率 深度可分离卷积 多尺度交叉模块 级联 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]