Evaluation of Image Inpainting Methods for Face Reconstruction of Masked Faces

In the field of facial recognition, recognizing masked faces presents a persistent challenge, with significant implications for various industries. Image inpainting emerges as a potential solution, aiding in the reconstruction of obscured facial features. Typically used to restore old or damaged images or remove unwanted elements, inpainting's application to masked faces is notably complex. In this study, we compare three cutting-edge image inpainting models: PatchMatch, Edge Connect, and Free-form image inpainting with gated convolution, assessing their performance in reconstructing masked faces. While these models excel at recreating natural scenes and objects, generating visually convincing unmasked faces from obscured images proves demanding. Our evaluation is based on two synthetic datasets we curated, MaskedFace-CelebA and MaskedFace-CelebA-HQ, containing masked facial images. Notably, the Gated Convolution model outperforms the other two in reconstructing facial images compared to their original counterparts. Furthermore, we employ a simple feed-forward neural network, the Extreme Learning Machine(ELM) classifier, to gauge the classification accuracy of reconstructed and ground truth images. The classifier's results substantiate both quantitative and qualitative image quality assessments.