胶囊网络对复杂现实场景中的物体识别

Capsule Network for Object Recognition in Complex Real Scenes

为了提高复杂场景目标图像的识别准确率,本文基于胶囊网络中的向量神经元的思想,在CapsNet网络基础上提出了一种改进的胶囊网络,用于实现复杂现实场景中的物体识别。改进的胶囊网络由两个卷积层和三个具有不同维度的胶囊层构成,在CapsNet网络结构的基础上进行了优化,在其CapsNet初级胶囊层之前增加了一层卷积层,并且在网络识别结构的后半部分增加了过滤胶囊层。该网络胶囊层中低层特征利用姿态关系对高层特征进行了预测,并采用动态路由算法和筛分决策机制最终选择性激活高级特征胶囊结构。实验结果表明,相较于CapsNet网络,文中网络对于同一复杂场景下目标图像的识别准确率提高了3.2%,且重构效果也较CapsNet有所提升,降低了复杂场景对于识别物体的干扰,提高了物体表征能力。

In order to improve the accuracy of target image recognition in a complex scene,the paper presents an improved capsule network based on the idea of vector neurons in capsule networks to identify the objects in complex real scenes.The improved capsule network consists of two convolutional layers and three capsule layers with different dimensions.Based on the CapsNet network,the network is optimized,with a layer of convolution added before the primary capsule layer and a filter capsule layer added to the second half of the network identification structure.The low-level features in the capsule layer use the attitude relationship to predict the high-level features and use the dynamic routing algorithm and the screening decision mechanism to selectively activate the advanced features.Experimental results show that,compared with the CapsNet network,the recognition accuracy of the improved network is improved by 3.2%,and the reconstruction effect has also been improved.It is more suitable for image processing with complex local relations.