基于深度可分离卷积的交叉模块的超分辨率重构算法

Super-Resolution Reconstruction Algorithm for Cross-Module Based on Depth-Wise Separable Convolution

摘要: 深度卷积神经网络进行单图像超分辨率重构方面已取得显着成果。大多数模型采用单流结构，不利于信息充分流动。采用交叉模块进行改善信息流动以获取足够的细节信息，然该方法虽然可以提高网络重构性能，但增加大量的网络训练参数，增大了网络训练难度，因此，本文提出了基于深度可分离网络的交叉模块(Cross-Module Based on Depth-Wise Separable Convolution, CM-DWSC)超分辨率重构算法，在交叉模块中使用深度可分离卷积代替普通卷积，同时，去掉与卷积层消耗等量内存的BN层，极大降低计算复杂度及模型容量。在每个级联子网络中堆叠多个改进交叉模块以便融合互补信息，有利于信息间流动，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，本文方法能够在减少网络参数情况下优于主流的超分辨率方法。

Abstract: The deep convolutional neural network has achieved remarkable results in single-image su-per-resolution reconstruction. Most models use a single-flow structure, and it is difficult to the full flow of information. The cross-module is used to improve the information flow to obtain enough detailed information. Although this method can improve the network reconstruction performance, it increases parameters and difficulty of network training. Therefore, this paper proposes the cross module base on depth-wise separable convolutional (CM-DWSC) algorithm for super-resolution reconstruction. Using depth separable convolution instead of ordinary convolution in the cross module, and the BN layer that consumes the same amount of memory as the convolutional layer is removed, greatly reduces the computational complexity and model capacity. A series of improved cross-modules are stacked in each cascade sub-network to fuse complementary information, which facilitates the flow of information. We introduce a residual learning strategy at each stage to utilize fully low-resolution feature information to further improve the reconstruction performance. The evaluation of the benchmark data set shows that the proposed method is superior to the mainstream super-resolution method in the case of reducing network parameters.