Wavelet scattering is a widely used feature extraction method due to its efficacy in extracting invariant features. Different wavelet scattering transforms exist for a varying number of input dimensions utilising various filter bank construction techniques. In this work, we propose a new scattering transform by generalising the 2-D joint time-frequency scattering filter bank to general 2-D and 3-D applications. Separable filters form the basis of joint 2-D and 3-D scattering, allowing for computational advantages in deep learning and more flexibility in filter bank construction compared to conventional multidimensional transforms. We demonstrate that the proposed scattering transform is nearly as effective as conventional 2-D scattering with MNIST handwritten digit classification and is more effective than 3-D scattering for hyperspectral image classification using a simple linear classifier. We also achieve the best results compared to benchmark methods on three of six MedMNIST3D voxel datasets using a simple neural network classifier on scattering features.
Rademan et al. (Thu,) studied this question.