Abstract
GAN-based image inpainting provides an effective framework for reconstructing missing image regions. However, the influence of mask representation, particularly mask color, on GAN-based reconstruction performance remains insufficiently investigated. This study systematically evaluates the effect of mask color on the reconstruction performance of the Pix2Pix conditional GAN for Khmer architectural ornament images. A dataset of 2,500 images acquired from Banteay Srei Temple was divided into 80% training and 20% testing images, with all images resized to 256 × 256 pixels. A fixed square mask covering 20% of the image area and positioned at the center was applied using five RGB color configurations: black (0,0,0), blue (0,0,255), gray (128,128,128), red (255,0,0), and white (255,255,255). The original Pix2Pix architecture was trained independently for each mask-color condition for 200 epochs with a batch size of one. Reconstruction performance was quantitatively assessed using Structural Similarity Index Measure (SSIM), Multi-Scale SSIM (MS-SSIM), Mean Squared Error (MSE), and Learned Perceptual Image Patch Similarity (LPIPS). The results demonstrate measurable differences in reconstruction performance across mask-color conditions. Blue masks achieved the highest SSIM (0.78399) and MS-SSIM (0.72768), whereas gray masks achieved the lowest MSE (0.01159) and black masks achieved the lowest LPIPS (0.15038). White masks exhibited comparatively lower performance, particularly in SSIM and LPIPS. These findings demonstrate that mask color is a relevant preprocessing factor in Pix2Pix-based image reconstruction and that its influence varies across structural, pixel-level, and perceptual quality measures. The study provides an empirical basis for selecting mask representations in deep learning-based reconstruction of cultural heritage imagery.
Keywords: Khmer architectural heritage; image reconstruction; image inpainting; Pix2Pix; mask color; deep learning; cultural heritage preservation.
About the presenter
Chanarin HENG is a PhD student in the Graduate School at Royal University of Phnom Penh, Cambodia. His research focuses on deep learning, computer vision, image inpainting, and digital cultural heritage preservation. His current work investigates artificial intelligence techniques for restoring damaged Khmer architectural ornaments to support the documentation and conservation of Cambodia's cultural heritage. He has published research on image restoration and deep learning applications in Khmer font style transfer.
Join us at the Forum, 5–6 November 2026, Institute of Technology of Cambodia.
Are you the author? Sign in to My NRF to share this abstract.