1: \begin{abstract}

2:     Recently, non-convex regularisation models have been introduced

3:     in order to provide a better prior for gradient distributions 

4:     in real images. They are based on using

5:     concave energies $\phi$ in the total variation type functional

6:     $\TV^\phi(u) \defeq \int \phi(|\nabla u(x)|) \d x$.

7:     In this paper, it is demonstrated that for typical choices of $\phi$, functionals

8:     of this type pose several difficulties when extended to the entire

9:     space of functions of bounded variation, $\BVspace(\Omega)$. 

10:     In particular, if $\phi(t)=t^q$ for $q \in (0, 1)$ and $\TV^\phi$ 

11:     is defined directly for piecewise constant functions

12:     and extended via weak* lower semicontinuous envelopes

13:     to $\BVspace(\Omega)$, then still $\TV^\phi(u)=\infty$

14:     for $u$ not piecewise constant. If, on the other hand,

15:     $\TV^\phi$ is defined analogously via continuously 

16:     differentiable functions, then $\TV^\phi \equiv 0$, (!).

17:     We study a way to remedy the models through

18:     additional multiscale regularisation and area strict

19:     convergence, provided that the energy $\phi(t)=t^q$ is 

20:     linearised for high values. The fact, that this kind of energies

21:     actually better matches reality and improves reconstructions,

22:     is demonstrated by statistics and numerical experiments.

23:     

24:     \paragraph{Mathematics subject classification:}

25:         26B30,  % Absolutely continuous functions, functions of bounded variation

26:         49Q20,  % Variational problems in a geometric measure-theoretic setting

27:         65J20.  % Improperly posed problems; regularization. (65Jxx Numerical analysis in abstract spaces)

28: 

29:     \paragraph{Keywords:} 

30:         total variation,

31:         non-convex,

32:         regularisation,

33:         area-strict convergence,

34:         multiscale analysis.

35:     

36: \end{abstract}

37: