On the applicability of speckle pattern imaging combined with AI for raw milk classification

This work demonstrates the high potential of an innovative technique for raw milk classification based on the integration of speckle pattern imaging and artificial intelligence. By exciting speckle patterns with a semiconductor laser and collecting experimental images with a CMOS camera, a total of 20 samples of raw cow milk with similar nutritional contents were tested during 4 Campaigns. Data analysis was conducted leveraging one common feature-based machine learning model and one state-of-the-art image-based deep learning model for speckle patterns. This study aims to provide in-depth insights to the community on how this measurement technique can be applied to raw cow milk samples and how the prediction models tested perform due to the similarity of the nutritional components of the samples. The machine learning model was trained on a set of 16 custom features, while the deep learning model used speckle pattern images as input. Both types of data were standardized dataset-wise beforehand using z-score. The best machine learning and deep learning models achieved 95% accuracy. The study highlights that the nutritional similarity of the samples highly impacts the models’ confusion in both cases, especially when Campaigns conducted at different sample temperatures were not included in the training. Overall, the analysis technique presented leveraging uncertainty metrics is a stepping stone toward relevant advances in the field of milk analysis.