Neural Network Texture Prediction Using Optical Properties
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for predicting visual texture parameters of effect coatings, such as those used in the automotive industry, require extensive training data and are cumbersome to adapt when new colorants are added, due to their reliance on specific concentrations of color components.
Innovation Solution
A neural network based on backpropagation that uses optical properties and parameters derived from physical models, rather than specific color component concentrations, to predict visual texture parameters, allowing for a more generalized and less complex training process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If regression-based methods or traditional neural networks are used to predict visual texture parameters based on colorant concentrations, then prediction capability is achieved, but the system requires extensive training data and is cumbersome to adapt when new colorants are added
Solution Approach 1:
The patent transforms the input parameters from specific colorant concentrations to optical properties (absorption coefficient K, scattering coefficient S, spectral reflectance) that are derived from physical models. This parameter transformation allows the neural network to generalize across different colorants and formulations without requiring retraining, as optical properties capture the essential light-interaction characteristics that determine visual texture regardless of the specific colorant composition
Solution Approach 2:
The patent introduces optical properties as intermediary parameters between the paint formulation and the visual texture parameters. Instead of directly mapping colorant concentrations to texture parameters, the system first calculates optical properties (K, S, spectral reflectance) from the formulation using physical models, then uses these optical properties as inputs to the neural network. This intermediary layer decouples the prediction system from specific colorant formulations, enabling adaptability to new colorants without extensive retraining
2Measurement precision
If neural networks are trained on specific color component concentrations, then texture parameter prediction is achieved, but the device complexity and training data requirements increase significantly
Solution Approach 1:
The patent reduces input dimensionality by changing from using individual colorant concentrations (which can number in the dozens or hundreds) to using aggregated optical properties (absorption coefficient K, scattering coefficient S, and spectral reflectance). This parameter aggregation dramatically reduces the number of input neurons required in the neural network while preserving the essential information needed for accurate texture parameter prediction
Solution Approach 2:
The patent replaces the need for extensive empirical training data collection with a physics-based approach. Instead of requiring numerous measured samples to train the network on colorant concentration relationships, the system uses physical models (Kubelka-Munk theory, Hapke model) to calculate optical properties from formulation data. This substitution of physical modeling for empirical data collection reduces training complexity and data requirements
Data Source
Figure 1
Figure 2
AI summary
The invention relates to a method for predicting visual texture parameters (9) of a paint with a known paint formulation (1). On the basis of a number of color components used in the paint formulation (1), visual texture parameters (9) of the paint are ascertained using an artificial neural network (7), and a value of at least one characteristic (5) which describes at least one optical property is ascertained for the known paint formulation (1) using a physical model (3), assigned to the known paint formulation (1), and transmitted to the artificial neural network (7) as an input signal in order to ascertain the visual texture parameters (9). The ascertained value assigned to the known paint formulation describes the at least one optical property for at least some of the number of color components of the paint formulation (1). In order to train the neural network (7), color samples with a respective known paint formulation are used, and for each color sample, the respective visual texture parameters are measured and assigned to a value, which is determined for the corresponding respective paint formulation, of the at least one characteristic that describes the at least one optical property for the respective paint formulation.