Background Object Removal and Image Inpainting to Fill Irregular Holes

Abstract

In the current scenario, nearly all deep learning methods to perform image inpainting use a standard convolution network in order to regenerate the holes created by removing the object. However, the result of this procedure is not up to the mark as the resultant image often appears to be distorted and blurry. Post-processing techniques do address this issue but prove to be computationally expensive. We aim to remove the background object from the image while ensuring that the regenerated image has minimal discrepancies. In our approach, we make use of partial convolution in which at every stage of the convolution network, we replace the convolution block with a partial convolution block. We have also focused on producing a high-quality mask dataset which ensures that our model works well on real life images where aberrantly shaped and sized objects are needed to be removed. Qualitative comparisons help analyze the predictions of our model in relation to those made by the standard image inpainting model of OpenCV.