1: \begin{abstract}

2: In this paper, we study the problem of a batch of linearly correlated image alignment,

3: where the observed images are deformed by some

4: unknown domain transformations, and  corrupted

5: by additive Gaussian noise and sparse noise simultaneously.

6: By stacking these images as the frontal slices of a third-order tensor,

7: we propose to utilize the tensor factorization method via transformed tensor-tensor product to

8: explore the low-rankness of the underlying tensor,

9: which is factorized into the product of two smaller tensors via transformed tensor-tensor product

10: under any unitary transformation.

11: The main advantage of transformed tensor-tensor product is that

12: its  computational complexity is

13: lower compared

14: with the existing literature based on transformed tensor nuclear norm.

15: Moreover, the tensor $\ell_p$ $(0<p<1)$ norm is employed to characterize

16: the sparsity of sparse noise

17: and the tensor  Frobenius norm is  adopted to model additive Gaussian noise.

18: A generalized Gauss-Newton algorithm is designed to solve the resulting model

19:  by linearizing the domain transformations

20:  and a proximal

21: Gauss-Seidel algorithm is developed to solve the corresponding subproblem.

22: Furthermore, the convergence of

23: the proximal Gauss-Seidel algorithm is established, whose

24: convergence rate  is also analyzed based on the Kurdyka-$\L$ojasiewicz property.

25: Extensive numerical experiments on real-world image datasets

26: are carried out to demonstrate the superior performance of

27: the proposed method  as compared to  several state-of-the-art methods

28: in both accuracy and computational time.

29: \end{abstract}

30: