Deep neural network optimization for pigment classification: a case study on representative medieval illumination mock-ups.

Neural networks are increasingly used in conservation science to exploit hyperspectral imaging (HSI) datasets. The literature includes a wide range of models with different architectures and hyperparameters available to the community. Yet, defining the optimal configurations of such parameters for new datasets is not straightforward. This work focuses on state-of-the-art optimization strategies to ensure that the best performance of the models is achieved, offering new opportunities for improving the predictive capacity of the model on unseen data. In particular, a synthetic mixture dataset is designed to train two widely-used neural network models: a 1D-CNN and a fully connected deep neural network (FC-DNN). The two models with architecture and hyperparameters taken directly from the literature are compared with optimized models using Bayesian approaches that combine Neural Architecture Search (NAS) and hyperparameter optimization (HPO). Full optimization of both the architecture and the hyperparameters of the neural network have proven to improve significantly the pigment classification accuracy, enabling more reliable pigment mapping from reflectance imaging spectroscopy (RIS) data. Optimized 1D-CNN and FC-DNN present similar capabilities for multilabel pigment classification of artistic materials, as shown through a representative case study on medieval illumination mock-ups.