Combining computer vision with the convolutional neural network to quantitatively analyse of the astaxanthin isomers in Antarctic krill meal

Abstract In this study, computer vision and deep learning was combined to develop a rapid method for quantifying the astaxanthin isomer content in krill meal. A total of 310 Antarctic krill meal samples were collected and their astaxanthin isomer content was determined as observed values using high-performance liquid chromatography. A computer vision system was then used to acquire images of the krill meal samples, which were subsequently preprocessed and fed into a Convolutional Neural Network (CNN) to establish a predictive model; its performance was compared with that of a feature-based artificial neural networks model. The results showed that the 13-cistrine (13-Cis) astaxanthin, all-trans astaxanthin, and 9-cistrine (9-Cis) astaxanthin content were distributed in the range of 0–2.05 mg/kg, 0.09–62.97 mg/kg, and 0–7.58 mg/kg, respectively. For the test set, CNN achieved an R2 of 0.96 in predicting all-trans astaxanthin and an R2 of 0.88 for 9-Cis astaxanthin. In out-of-sample validation, the CNN achieved mean relative errors of 5.20% and 11.35% for all-trans and 9-Cis astaxanthin, respectively. In conclusion, computer vision combined with CNN offers an efficient, precise, and non-destructive technique for quantitatively analysing astaxanthin isomers in krill meal.