Recurrent feature reasoning network for image inpainting : a comprehensive study and analysis

Abstract

This study replicates and examines the Recurrent Feature Reasoning (RFR) network for image inpainting, originally proposed for repairing large continuous holes in images by inferring hole boundaries. The replication utilizes only a small subset of the Paris dataset, allowing me to assess the functionality of the network within the constraints of low-dimensional data. Nevertheless, despite including relatively few inputs, the implementation follows the same methodology proposed in the original RFR article, in that it iterates over the feature maps to improve the inpainting results. The results of the replication indicate that, even when limited with regards to the amount of available data, RFR-Net continues to function and inpaint in an effective manner. Thus, the results suggest that the framework of RFR-Net is relatively solid and extends its potential application. This can contribute to a better understanding of the framework’s limitations in the case of limited data and allow to assess its functionality. This replication, as such, also helps clear the pathway for further research in the inpainting of contextual images blind with low quantity and quality of data.