Multi spectral imaging analysis for meat spoilage discrimination

In the present study, fresh beef fillets were purchased from a local butcher shop and stored aerobically and in modified atmosphere packaging (MAP, CO2 40%/O2 30%/N2 30%) at six different temperatures (0, 4, 8, 12, 16 and 20°C). Microbiological analysis in terms of total viable counts (TVC) was performed in parallel with videometer image snapshots and sensory analysis.

Odour and colour characteristics of meat were determined by a test panel and attributed into three pre-characterized quality classes, namely Fresh; Semi Fresh and Spoiled during the days of its shelf life. So far, different microbiological and (bio)chemical methods are employed to assess meat spoilage, the majority of which are slow, time-consuming and expensive procedures and thus, it would be most preferable to be replaced by faster and directly applicable methods.

Therefore developing a procedure by associating image data with corresponding sensory data would be of great interest. The purpose of this research was to produce a method capable of quantifying and/or predicting the spoilage status (e.g. express in TVC counts as well as on sensory evaluation) using a multi spectral image of a meat sample and thereby avoid any time-consuming microbiological tests.

To accomplish this, first the images were converted into values that were comparable to the corresponding data, using the Minimum Noise Fraction (MNF) transformation and simple thresholding. Moreover, association of image data with sensory data was undergone using three different classification methods: Naive Bayes Classifier as a reference model, Canonical Discriminant Analysis (CDA) and Support Vector Classification (SVC).

As the final step, generalization of the models was performed using k-fold validation (k=10). Results showed that image analysis provided good discrimination of meat samples regarding the spoilage process as evaluated from sensory as well as from microbiological data. The support vector classification (SVC) model outperformed other models.

Specifically, the misclassification error rate (MER), derived from odour characteristics, was 18% for both aerobic and MAP meat samples. In the case where all data were taken together the misclassification error amounted to 16%. When spoilage status was based on visual sensory data, the model produced a MER of 22% for the combined dataset.

These results suggest that it is feasible to employ a multi spectral image for the quantitative determination of meat spoilage status during storage in different conditions.