AI 용어집
인공지능 완전 사전
Automatic Colorization
Process of predicting color information (chrominance channels) from grayscale images (luminance) using deep neural networks.
Pretext Task
Auxiliary objective designed to enable self-supervised learning, such as colorization, which forces the model to learn useful semantic representations.
LAB Color Space
Colorimetric space separating luminance (L) from chrominance components (A and B), commonly used for colorization as it allows working independently on brightness and colors.
Perceptual Loss
Loss function measuring perceptual similarity between images by comparing their high-level representations rather than individual pixels, improving the visual quality of results.
GAN for Colorization
Generative Adversarial Network applied to colorization, where a generator produces colorized images and a discriminator evaluates their authenticity against real color images.
Color Quantization
Process of discretizing continuous color space into discrete classes, transforming the colorization problem into a more stable multi-class classification task.
Temperature Rescaling
Technique applied to color probability distributions to control the diversity of predicted colors, avoiding desaturated results and encouraging more vibrant predictions.
Multi-scale Colorization
Approach processing the image simultaneously at multiple resolutions, allowing capture of both global structure and fine details for more consistent and realistic colorization.
Attention mechanism
Module allowing the model to selectively weight different regions of the image during color prediction, improving the semantic coherence of results.
Unsupervised learning
Ability to learn from unannotated data by exploiting the intrinsic structure of data, such as the correlation between luminance and chrominance in natural images.
Reconstruction loss
Objective function measuring the discrepancy between predicted and actual colors, often combined with other losses to balance accuracy and perceptual realism.
Conditional color histogram
Statistical representation of color distribution in an image, used as an additional condition to guide the colorization process toward specific palettes.