A novel approach for recipe prediction of fabric dyeing based on feature-weighted support vector regression and particle swarm optimisation

Recipe prediction is one of the most critical steps in the fabric dyeing industry. The conventional Kubelka-Munk model and neural network techniques have been widely used in recipe prediction systems. However, there are some limitations to these two methods: predictions using the Kubelka-Munk model may not be robust enough; and neural networks require large amounts of training data. Therefore, this paper investigates a novel recipe prediction method for fabric dyeing based on feature-weighted support vector regression and particle swarm optimisation. Feature-weighted support vector regression improved with particle swarm optimisation was first developed to predict the CIELab coordinates for given dye concentrations, expressed as (c(1), c(2),...,c(n)) double right arrow (L*, a*, b*). Particle swarm optimisation was utilised again in the recipe prediction stage to search for the optimal recipe in an iterative process. The optimisation criterion is to minimise the colour differences (CMC [2:1]) between the CIELab value calculated by feature-weighted support vector regression improved with particle swarm optimisation and the target CIELab of a swatch. Dyeing data based on two different fabrics (cotton and taffeta) were used in the experiment. The proposed method revealed good results with a slight average colour difference between the target and reproduced colours. The absolute percentage errors in predicting concentrations were less than 5% in most of our experimental recipes. In addition, the comparative experimental results illustrate that our method had higher accuracy and better practical applicability than other methods.