Predicting Color Schemes Based on Genetic Algorithm and Deep Feedforward Neural Network Model

To enhance the accuracy and efficiency of color matching for opaque products, a relevant model is employed to predict their color schemes. By extracting the spectral reflectance of dyed panels as input and dyeing ratios as output, it is possible to reduce color matching errors and improve production efficiency, thereby advancing the dyeing technology of opaque products and achieving color matching freedom. This study initially applies a logarithmic transformation to the K/S (absorption coefficient/scattering coefficient) ratio and concentration, converting the resultant equations into linear forms. The least squares method is then employed to solve the linear equations describing the relationship between K/S and concentration. A comparison is made between linear regression fitting and nonlinear regression fitting methods, revealing that the former yields better fitting results and higher correlation coefficients in the range of 400nm-700nm for K/S versus concentration. Thus, linear fitting is chosen. In addressing the issue of predicting color differences with minimal deviation based on target samples, this paper unifies variables using the Beer-Lambert color matching equation, Kubelka-Munk theory, and the Simpson discrete variable integration formula to obtain a training set. Leveraging the CIELAB color space model, a deep feedforward neural network is trained on the L, a, and b values for red, yellow, and blue. Finally, a predictive model combining genetic algorithm and deep feedforward neural network (GA-DFNN) is utilized to forecast color matching schemes for the original samples, resulting in 5 sets of color matching schemes with color differences of no more than 1 unit.