Prediction of Yogurt Quality Attributes from Hyperspectral Images using Deep Learning

Food quality attributes such as pH, lactic acid concentration, and viscosity are important in evaluating the stability, safety, and textural properties of refrigerated dairy products. This study uses yogurt as a test case for predicting these attributes from hyperspectral images. The proposed method uses raw spectral data collected from yogurt samples made using two different milk sources: cow milk and coconut milk. We developed an automated pipeline for predicting seven food quality attributes of yogurt, including pH, lactic acid concentration, viscosity, and color-based quality values. The pipeline uses a stepwise segmentation method to select the yogurt region, which is then divided into fixed-size spectral patches. Each patch contains pixels from only the yogurt region. We designed a convolutional neural network which is trained using these patches along with temporal information. The network includes a shared encoder and separate prediction heads for the two different milk sources. Our evaluation shows high test-set $R^{2}$ values (above 0.85) for $\mathbf{p H}$ and color attributes ($\mathrm{l}, \mathrm{a}, \mathrm{b}$), and moderate values for titratable acidity (0.43) and viscosity (0.68). This result indicates that non-invasive hyperspectral imaging provides reliable prediction for key yogurt quality attributes, especially when combined with temporal information.