Scattering Transform and Sparse Linear Classifiers for Art Authentication

Roberto Leonarduzzi, Haixia Liu, Yang Wang

Abstract:

Recently, a novel signal-processing tool was proposed, the scattering transform, which uses a cascade of
wavelet filters and nonlinear (modulus) operations to build translation-invariant and deformation-stable
representations. Despite being aimed at providing a theoretical understanding of deep neural networks, it
also shows state-of-the-art performance in image classification. In this paper, we explore its performance
for art authentication purposes. We analyze two databases of art objects (postimpressionist paintings and
Renaissance drawings) with the goal of determining those authored by van Gogh and Raphael, respectively. To that end, we combine scattering coefficients with several linear classifiers, in particular sparse $\ell_1$-regularized classifiers. Results show that these tools provide excellent performance, superior to state-of-the-art results. Further, they suggest the benefits of using sparse classifiers in combination with deep networks