Capsule networks overcome the two drawbacks of convolutional neural networks: weak rotated object recognition and poor spatial discrimination. However, they still have encountered problems with complex images, including high computational cost and limited accuracy. To address these challenges, this work has developed effective solutions. Specifically, a novel windowed dynamic up-and-down attention routing process is first introduced, which can effectively reduce the computational complexity from quadratic to linear order. A novel deconvolution-based decoder is also used to further reduce the computational complexity. Then, a novel LayerNorm strategy is used to pre-process neuron values in the squash function. This prevents saturation and mitigates the gradient vanishing problem. In addition, a novel gradient-friendly network structure is developed to facilitate the extraction of complex features with deeper networks. Experiments show that our methods are effective and competitive, outperforming existing techniques.

胶囊网络克服了卷积神经网络的两个缺点：旋转对象识别能力弱和空间辨别力差。然而，他们仍然遇到了复杂图像的问题，包括计算成本高和准确性有限。为了应对这些挑战，这项工作开发了有效的解决方案。具体来说，首先引入了一种新的窗口式动态上下注意力路由过程，可以有效地降低从二次级到线性级的计算复杂度。还使用了一种新颖的基于反卷积的解码器来进一步降低计算复杂性。然后，使用一种新的 LayerNorm 策略对 squash 函数中的神经元值进行预处理。这可以防止饱和并缓解梯度消失问题。此外，还开发了一种新的梯度友好型网络结构，以促进提取具有更深网络的复杂特征。实验表明，我们的方法有效且具有竞争力，优于现有技术。