Recoloração de Obras de Arte Digitalizadas para Pessoas com Daltonismo Utilizando Redes Neurais Convolucionais

Abstract

Color blindness is a visual deficiency that impairs color distinction, limiting the aesthetic and semantic experience of artworks for millions of individuals. This work developed and validated a computer vision model for the automatic recoloring of digitized artworks, aiming at visual inclusion for individuals with protanopia, deuteranopia, tritanopia, and monochromacy. Unlike approaches based on adversarial networks, the adopted methodology consisted of a convolutional neural network with an autoencoder architecture and skip connections, trained to minimize a perceptual loss function based on the VGG19 network. The training utilized the WikiArt dataset, processing an objective function that approximates the simulated view of the recolored image to the original image. Quantitative results demonstrated the effectiveness of the method, achieving an overall average of 0.94 in the Structural Similarity Index (SSIM) and 21.83 dB in the Peak Signal-to-Noise Ratio (PSNR). The comparative visual assessment confirmed that the model preserved the structural integrity of the artworks while restoring chromatic contrast in conflicting regions, validating the use of autoencoders guided by perceptual loss as a robust and efficient solution for accessibility in digital collections.