Prediction of translucent plastic colour formulation based on Transformer and parallel deep convolutional neural network

Abstract Colour formulation prediction based on a neural network aims to achieve target colours through various colourants. Currently existing methods cannot fully fit the non‐linear relationship between target colour and formulation because of neglect of the potential regularities between colour samples from different processes and the lack of effective utilisation of global features and local features. To address these issues, a novel method of colour information combination and the Global‐Detail Colour Feature Fusion (GD‐CFFusion) network framework is proposed for the first time. Complete formulation information under specific process conditions can be provided by colour information combinations, which consist of reflectance curves of target colours, CIELab values, colourant selections, contrast considerations, backing types and colourant concentration gradients. GD‐CFFusion consists of a global feature enhancer based on Transformer, a detail feature enhancer based on a convolutional neural network and a feature fusion summariser. It achieves the precise fitting of the non‐linear relationships between colour information combinations and formulations as well as reflectance curves of formulation. Additionally, the underlying patterns between colour samples from different processes and intrinsic relationships of each sample through colour information combination can be learned. The plastic colour matching dataset consists of collated plastic colour samples on which GD‐CFFusion was trained. Experimental results showed that GD‐CFFusion achieved a total loss of 0.0203 and a formulation loss of 0.0388 on the test set of 4114 samples. This result outperforms all baseline networks and most other colour formulation prediction methods.