When dissolving dairy-based powder in water, some particles do not ‘instantly’ dissolve and remain unreconstituted. Current tests for inspecting these particles are mainly prone to subjective interpretation and may not be reproducible. A reproducible quantitative detection can improve the quality assurance process and may permit a more accurate correlation of the defect with its underlying causes. In this study, a collaborative robot rehydrated twenty-nine commercial and pilot stage-1 infant formula powders and photographed the mixtures in a commercially available baby bottle. Although image processing using deep learning techniques is popular when large data sets are available, this study, which had a relatively small dataset, aimed to evaluate the efficacy of classical image processing techniques based on mean shift and sharpness. To compare our proposed method with commonly employed methods, a human rated the images using a number of developed reference pictures. In this preliminary evaluation, a Spearman correlation of 0.68 was observed between computer vision and human ratings for the same images. Further work is planned to evaluate the method, but this result suggests that the proposed method can detect and quantify undissolved particles in reconstituted stage-1 infant formula in a commercially available baby bottle, comparable to an end-user evaluation.