一种改进的Capsule及其在SAR图像目标识别中的应用

为了解决Capsule网络随着输入图像增大计算量和参数数量急剧增加的问题，对Capsule网络进行了改进并将其用于SAR自动目标识别（SAR-ATR）中。基于大脑视觉皮层以层级结构以及柱状形式处理信息的机制，提出了完全实例化的思想，并运用类脑计算对Capsule网络进行了改进。具体方法是:使用多个卷积层实现层级处理，同时使用了较少的卷积核，但每一层使用的卷积核数量随着层级加深逐渐增加，使得提取的特征更加趋于抽象化；在PrimaryCaps层中，Capsule向量由最后一层卷积层输出的所有特征图构成，使得Capsule单元包含目标局部或整体的全部特征，以实现目标的完全实例化。在SAR-ATR上，将改进的Capsule网络与原Capsule网络、传统目标识别算法和基于经典卷积神经网络的目标识别算法进行对比实验。实验结果表明，改进的Capsule网络训练参数和计算量大大减少，并且训练速度得到很大提升，在SAR图像数据集上的识别准确率较Capsule网络和前两类方法分别提高了0.37和1.96~8.96个百分点。

An improved Capsule and its application in target recognition of SAR images

In order to solve the problem that the Capsule network increases the amount of calculation and the number of parameters increases sharply with the input picture, the Capsule network is improved and the improved Capsule network is used in SAR automatic target recognition (SAR-ATR). In this paper, based on the mechanism of brain visual cortex processing information in hierarchical structure and column form, the idea of complete instantiation was proposed, and the brain-like calculation was used to improve the Capsule network. The specific method was to use multiple convolution layers to achieve hierarchical processing. The number of convolution kernels used in each layer increases with the depth of the hierarchy, which made the extracted abstract features gradually increase. In the PrimaryCaps layer, the Capsule vector consisted of all the feature maps output by the last layer of the convolutional layer, so that the Capsule unit contained all the features of the target part or the whole to achieve full instantiation of the target. On the SAR-ATR, a comparison experiment was performed with the Capsule network, the traditional target recognition algorithm and the target recognition algorithm based on the classical convolutional neural network. The experimental results show that the improved Capsule network training parameters and calculations are greatly reduced, and the training speed is greatly improved, and the recognition accuracy on the SAR image data set is increased by 0.37 and 1.96-8.96 percentage points compared with the Capsule network and the first two methods respectively.