Text2Color Networks: Deep Learning Models for Color Generation from Compositional Color Descriptions

Abstract

Color serves as an important cue in graphics, arts and in many computer vision applications. We reliably and effortlessly name colors and communicate them with their names. However, many applications such as graphics, color design, palette generation and color selection tools demand numerical values of colors. Predicting and communicating colors by their numerical values is less intuitive and difficult task as it is a mapping of millions of colors for a specific color space. To bridge the gap between linguistic color names and numerical color values, in this paper, we present neural network architectures that predict a point in color space for a given color name. Proposed models provide user interface between colors and names by learning the language semantics and mimic human level comprehension of color descriptions to predict colors and modify them with respect to linguistic adjectives of color names. We consider color prediction as a regression problem and solve it as a language modeling task. Color descriptions are taken as text sequences and each sentence is represented with word-level tokenization. Each token is transformed into a word vector in the latent space using CBOW word embeddings model. Word vectors representing color names are fed as input to neural networks and trained with normalized R, G, B values as supervision information. Trained models are capable of predicting color for a given color name and modify colors for different nouns and adjectives associated with color names. We also built color generation models based on pre-trained word embeddings to overcome the limited availability of large linguistic color name datasets. These pre-trained models perform well with datasets containing few thousand color names. We then present two recommendation engines that suggest similar color palette to user given color name. These recommendation engines enhance the color vocabulary and assist users in the color selection process.