A Scalable Spatial Anisotropic Interpolation Approach for Object Removal from Images Using Elastic Net Regularization

Abstract

Object removal from an image is a novel problem with a lot of applications, in the area of computer vision. The ill-posed nature of the problem and the non-stationary content present in the image render it a complicated task. The diffusion-based and self-similarity based algorithms available in the literature explicitly model either the structures or the textures but not the both. They are good at solving small instances of the problem. However, they tend to produce low fidelity results and turn out to be intractable if the relative size of the object to the input image increases. The moving average based Spatial Anisotropic Interpolation (SAI) for text removal, proposed in our previous work also failed due to its poor extrapolation capability. Thus, it is imperative to develop a sampling scheme which can retain the interpolation feature while showing an apposite concern to the non-stationary features present in the image. The proposed, Design of Computer Experiments (DACE) driven Scalable SAI (SSAI) is a natural extension of SAI in three aspects. Precisely, it extends the Systematic Sampling to ‘Not only Symmetric Hierarchical Sampling’ (NoSHS), intelligently selects a basis based on Hurst Exponent, and employs Elastic Net regularization of Gaussian regression error for determining the order of the polynomial. Hence, these adaptive features increase the fidelity of the results. This paper elaborates the proposed framework- SSAI and demonstrates its capabilities by comparing the results with the latest hybrid approaches using the PSNR metric.