Abstract:

Most of the current capsule network methods improve the classification accuracy by modifying iterative routing， while ignoring the burden brought by complex computation of iterative routing itself. Although there are some methods that use non-iterative routing to train the capsule network， the accuracies of these methods are not good. To address the above problem， a non-iterative routing graph capsule network method for remote sensing scene classification was proposed. Firstly， the preliminary features of the input image were extracted using a simple convolutional layer. Then， by performing dual attention between channels and capsules sequentially， a global attention module with dual fusion between channels and capsules was presented to generate global coefficients that weighed high-level capsule features. As a result， the weighted high-level capsule features became more discriminative to highlight the important capsules， thereby improving the classification performance. Meanwhile， an equivariant regularization term that could compute the similarity among the input images was introduced to model the explicit equivariance of the capsule network， thereby improving network performance potentially. Finally， the whole network was trained based on the loss function combining margin loss and equivariance loss to obtain a discriminative classification model. Experimental results on multiple benchmark scene datasets verified the effectiveness and efficiency of the proposed method. Experimental results show that the proposed method has the classification accuracy reached 90.38% on Canadian Institute For Advanced Research-10 image datasets （CIFAR-10）， which is 15.74 percentage points higher than the Dynamic Routing Capsule network （DRCaps） method， and achieves classification accuracy of 98.21% and 86.96% on Affine extended National Institute of Standards and Technology dataset （AffNIST） and Aerial Image Dataset （AID）， respectively. It can be seen that the proposed method can improve the performance of remote sensing scene classification effectively.

Key words: remote sensing scene classification, graph capsule network, non-iterative routing, equivariant regularization

摘要：

目前大多数胶囊网络方法通过改进迭代路由的方式提高分类精度，而忽略了迭代路由本身复杂的计算量带来的负担。虽然有方法采用无迭代的路由训练胶囊网络，但是精度不佳。针对以上问题，提出无迭代路由图胶囊网络的场景分类模型。首先，利用简单卷积层提取输入图像的初始特征；接着，提出通道和胶囊间双融合的全局注意力模块，通过依次进行通道和胶囊之间的注意力生成全局权重系数来加权高级胶囊特征，使加权后的高级胶囊特征更具判别性，以突出重要的胶囊，从而提高分类性能；同时，引入能计算图像间相似性的等变正则化项，以建模胶囊网络的显式等变性，从而潜在地提升网络性能；最后，基于边界损失和等变损失的组合损失函数训练整个网络，以得到富于判别性的分类模型。在多个基准场景数据集上的实验结果验证了所提方法的有效性和效率。实验结果表明，所提方法在加拿大高级研究所的10类图像数据集（CIFAR-10）上的分类准确率达到90.38%，与动态路由胶囊网络（DR-Caps）方法相比，提高了15.74个百分点；并且在仿射手写数字图像（AffNIST）数据集和航空影像数据集（AID）上，分别取得了98.21%和86.96%的分类准确率。可见，所提方法有效提高了场景分类性能。

关键词: 遥感场景分类, 图胶囊网络, 无迭代路由, 等变正则化