Blind image inpainting quality assessment using local features continuity

This paper deals with Blind Inpainted Image Quality Assessment BIIQA. Herein, we propose a new method that exploits the continuity of features around the boundaries of the retouched area. Indeed, we believe that the quality of inpainted images depends on how the edges and textures have been reproduced inside the hole. Besides, this concept has been formalized by the fact that features should be reproduced inside and outside the hole with respect to structures continuity. Furthermore, one could compare these features in terms of continuity and estimate the global quality of the inpainted image. And since the local structures are represented by patches, we proposed as a secondary contribution, an improvement of a patch classification algorithm. The strength of this metric unlike most existing IIQA metrics, is that it is completely blind and does not require any reference, making it well suited to the inpainting assessment, where reference is usually unavailable. The proposed BIIQA has been tested on TUM-IID database, where the results of four commonly used inpainting algorithms are provided and compared against IIQA state-of-the-art. The obtained results show clearly that our method outperforms the existing ones.