Parallel-Connected Residual Channel Attention Network for Remote Sensing Image Super-Resolution
Parallel-Connected Residual Channel Attention Network for Remote Sensing Image Super-Resolution
复制标题
DOI：
10.1007/978-3-030-69756-3_2
复制
发表时间：
2020
期刊：
影响因子：
--
通讯作者：
Yinhao Li;Yutaro Iwamoto;Lanfen Lin;Yenwei Chen
中科院分区：
文献类型：
--
作者：
Yinhao Li;Yutaro Iwamoto;Lanfen Lin;Yenwei Chen
近年来，卷积神经网络（CNN）在遥感图像的单图像超分辨率（SR）方面取得了可喜的成果。然而，大多数现有的方法是不充分的遥感图像SR由于所需的高计算成本。因此，提高表示能力，更少的参数和更短的预测时间是一个具有挑战性的和关键的任务，遥感图像SR。在本文中，我们提出了一种新的CNN称为并行连接的残差通道注意力网络（PCRCAN）。具体来说，受组卷积的启发，我们提出了一个并行模块与PCRCAN中的特征聚合模块。并行模块通过扩展网络架构，显著降低了模型参数，并完全集成了特征映射。此外，为了降低训练复杂深度网络的难度并提高模型性能，我们使用残差通道注意块作为基本特征映射单元，而不是单个卷积层。在公共遥感数据集UC默塞德土地利用数据集上的实验表明，PCRCAN比大多数最先进的方法具有更高的精度，效率和视觉改善。
In recent years, convolutional neural networks (CNNs) have obtained promising results in single-image super-resolution (SR) for remote sensing images. However, most existing methods are inadequate for remote sensing image SR due to the high computational cost required. Therefore, enhancing the representation ability with fewer parameters and a shorter prediction time is a challenging and critical task for remote sensing image SR. In this paper, we propose a novel CNN called a parallel-connected residual channel attention network (PCRCAN). Specifically, inspired by group convolution, we propose a parallel module with feature aggregation modules in PCRCAN. The parallel module significantly reduces the model parameters and fully integrates feature maps by widening the network architecture. In addition, to reduce the difficulty of training a complex deep network and improve model performance, we use a residual channel attention block as the basic feature mapping unit instead of a single convolutional layer. Experiments on a public remote sensing dataset UC Merced land-use dataset revealed that PCRCAN achieved higher accuracy, efficiency, and visual improvement than most state-of-the-art methods.