10 2. BASICS OF FEATURE DESIGN
A necessary condition for obtaining invariant scalar products is that the output space of the
operator f establishes a tight frame, i.e., the generalized Parseval’s identity applies. In certain
cases, this can only be achieved approximately due to effects of the pixel grid and deviations
from the invariance property can be used to optimize discrete operators on the grid.
For instance, the Scharr filter Scharr et al. [1997] has been optimized by minimizing the
anisotropy of the structure tensor. Note that different assumptions in the formulation of the
scalar product and thus the weights in the optimization lead to different results and the Scharr
filter might not be the most isotropic choice [Felsberg, 2011].
Further practical problems besides grid effects are caused by the final extent of image data.
Global scale-invariance can only be achieved if data is available on an infinite domain. Since
this is impossible in practice, the image domain has to be extended by other tricks, e.g., peri-
odic repetition (Fourier transform) or reflective boundaries (discrete cosine transform, DCT)
of a rectangular domain. However, these tricks hamper a proper rotation invariance. Using a
circular domain with reflective boundaries theoretically solves all issues, but becomes infeasible
to compute [Duits et al., 2003].
2.3 SPARSE REPRESENTATIONS, HISTOGRAMS, AND
SIGNATURES
Obviously, useful features need to be feasible to compute. Depending on the application, the
selection of features might be limited due to real-time constraints and this is actually one area
where deep features are still problematic. Also, the space complexity of features, i.e., their mem-
ory consumption, might be decisive for the design. In the past, paradigms have shifted regularly
back and forth between using compact features and sparse features [Granlund, 2000a].
e kernel trick in support vector machines (SVMs) and Gaussian processes are examples
of implicit high-dimensional spaces that are computationally dealt with in low-dimensional,
nonlinear domains. In contrast, channel representations and convolutional networks generate
explicit high-dimensional spaces. e community has conflicting opinions in this respect, but
recently, compactification of originally sparse and explicit features seems to be the most promis-
ing approach, also confirmed by findings on deep features [Danelljan et al., 2017].
Another strategy to improve computational feasibility and the memory footprint is to use
feedback loops in the feature extraction. Whereas deep features are typically feed-forward and
thus mostly do not exploit feedback, adaptive [Knutsson et al., 1983] or steerable [Freeman
and Adelson, 1991] filters are a well-established approach in designed feature extractors. In
relation to equivariance properties and factorized representations, adaptive filters often exploit
projections of the equivariant part of the representation, e.g., orientation vectors or structure
tensors.
Alternatively, iterative methods such as diffusion filtering can be applied [Weickert, 1996],
which potentially open up more efficient feature extraction using recurrent networks. e rela-
tionship between recurrent schemes and robust, unbiased feature extraction has been identified,