基于胶囊网络的三维模型识别

为解决传统卷积神经网络中大量池化层的引入导致特征信息丢失的问题,依据胶囊网络(CapsNet)使用向量神经元保存特征空间信息的特性,提出了一种用以识别三维模型的网络模型3DSPNCapsNet。使用新的网络结构,提取更具代表性的特征的同时降低了模型复杂度,并提出基于动态路由(DR)算法的DRL算法来优化胶囊权重的迭代计算过程。在ModelNet10上的实验结果表明,相比3DCapsNet以及VoxNet,该网络取得了更好的识别效果,在原始测试集上3DSPNCapsNet的平均识别准确率达到95%,同时验证了该网络对旋转三维模型的识别能力。适当扩展旋转训练集之后,所提网络对各角度旋转模型的平均识别率达到81%。实验结果表明,3DSPNCapsNet对三维模型及其旋转具有良好的识别能力。

3D model recognition based on capsule network

In order to solve the problem of feature information loss caused by the introduction of a large number of pooling layers in traditional convolutional neural networks, based on the feature of Capsule Network (CapsNet)——using vector neurons to save feature space information, a network model 3DSPNCapsNet (3D Small Pooling No dense Capsule Network) was proposed for recognizing 3D models. Using the new network structure, more representative features were extracted while the model complexity was reduced. And based on Dynamic Routing (DR) algorithm, Dynamic Routing-based algorithm with Length information (DRL) algorithm was proposed to optimize the iterative calculation process of capsule weights. Experimental results on ModelNet10 show that compared with 3DCapsNet (3D Capsule Network) and VoxNet, the proposed network achieves better recognition results, and has the average recognition accuracy on the original test set reached 95%. At the same time, the recognition ability of the network for the rotation 3D models was verified. After the rotation training set is appropriately extended, the average recognition rate of the proposed network for rotation models of different angles reaches 81%. The experimental results show that 3DSPNCapsNet has a good ability to recognize 3D models and their rotations.