Substantial computing costs are required to use deep-learning algorithms. Here, we implement feature extractionbased on analytic relations in the Fourier-transform domain. In an example relevant to visual odometry, wedemonstrate a reduction in algorithmic complexity with cross-power spectral preprocessors for feature extractionin lieu of learned convolutional lters. With spectral reparameterization and spectral pooling, not only can theoptical ow (spatial disparity of images in a sequence) be computed, but occluding objects can also be trackedin the foreground without deep learning. There is evidence that insects with small brains implement similarvisual-data spectral preprocessors, which may be critical in the development of future real-time machine learningapplications.
展开▼
