Multitask learning for denoising and analysis of X-ray polymer acquisitions

X-ray acquisitions are beneficial in food contaminant analysis as they can detect both metallic and non-metallic objects. This paper considers the scenario of single-pixel hyperspectral X-ray acquisitions applied to a series of materials with different characteristics. We propose a method that jointly applies a denoising operation and detects the analysed material in terms of a physical parameterisation. The proposed algorithm is based on a Convolutional Neural Network (CNN) trained with a multitask learning strategy using a custom loss function tailored to the problem at hand. Experimental results on metals and polymers show that the proposed method can also generalise to materials never seen at training time.