Prediction Model for Semitransparent Watercolor Pigment Mixtures Using Deep Learning with a Dataset of Transmittance and Reflectance (1904.00275v1)

Abstract: Learning color mixing is difficult for novice painters. In order to support novice painters in learning color mixing, we propose a prediction model for semitransparent pigment mixtures and use its prediction results to create a Smart Palette system. Such a system is constructed by first building a watercolor dataset with two types of color mixing data, indicated by transmittance and reflectance: incrementation of the same primary pigment and a mixture of two different pigments. Next, we apply the collected data to a deep neural network to train a model for predicting the results of semitransparent pigment mixtures. Finally, we constructed a Smart Palette that provides easily-followable instructions on mixing a target color with two primary pigments in real life: when users pick a pixel, an RGB color, from an image, the system returns its mixing recipe which indicates the two primary pigments being used and their quantities.

• The paper introduces a deep learning model (SWPM) and a novel dataset (NTU WPSM) using both transmittance and reflectance to accurately predict semitransparent watercolor pigment mixtures.
• The SWPM model demonstrated superior prediction accuracy compared to traditional Kubelka-Munk theory, achieving color differences ( ΔE *ab) below 5 in 83% of cases.
• The research led to the Smart Palette tool, which integrates the model to assist novice painters with precise color mixing, showing significant improvement in user color matching tasks.

Deep Learning-Based Prediction Model for Watercolor Pigment Mixtures

The paper presents a prediction model aimed at simplifying the understanding of color mixing for novice watercolor painters, addressing the complexity arising from semitransparent pigment mixtures. Utilizing deep learning techniques, the authors present the Smart Palette system, designed to provide targeted support for color selection by offering precise mixing recipes based on deep neural network (DNN) predictions.

In this paper, the researchers focus on developing a comprehensive dataset, NTU WPSM, which includes extensive transmittance and reflectance measurements for 13 primary pigments measured across 12 varying quantities. The unique aspect of this dataset is the dual parameters—reflectance and transmittance—that account for the semitransparent nature of watercolor pigments. The dataset is further enriched with measurements of mixed pigment combinations, encapsulating a complete spectrum of mixing possibilities crucial for a reliable prediction model.

The core of the paper is the SWPM prediction model, constructed via a DNN, which forecasts the results of mixing two primary pigments. The model's performance was validated against a ground truth with an accuracy, demonstrating that 83% of predictions yielded a color difference ($\Delta E^{*}_{ab}$) of less than 5, a threshold where average observers cannot distinguish between two colors. Moreover, when compared to traditional methods based on two-constant Kubelka-Munk (KM) theory, the prediction model showed enhanced accuracy. Twelve out of fifteen selected test cases confirmed a more precise color prediction using the proposed model, emphasizing its superiority over the classical approach.

In terms of application, the Smart Palette integrates the prediction model to guide users in color mixing. This tool leverages the look-up table created through extensive interpolation and prediction of color outcomes, allowing users to select and match colors from digital images with increased precision in the physical field. User studies involving 18 participants indicated a significant improvement in color matching tasks performed with the Smart Palette as compared to traditional intuition-based or Itten's color wheel methods.

The paper concludes with both theoretical and practical implications. Theoretically, it introduces a novel approach to the challenge of simulating semitransparent media mixing, offering potential expansion into other artistic media beyond watercolors. Practically, the Smart Palette tool positions itself as an effective educational aid for learners, potentially revolutionizing traditional practices in color education.

As a potential future direction, this research platform can be further extended to include a broader spectrum of pigments or adapted for other painting media such as acrylics or oils, which similarly involve complex optical interactions. Furthermore, addressing symmetry issues inherent in DNN predictions could enhance the model's reliability. These enhancements could culminate in a comprehensive, interdisciplinary tool applicable both within digital graphics and real-world artistic endeavors.