Automated identification and visualization of food defects using RGB imaging: Application to the detection of red skin defect of raw hams

Colourgrams are signals that codify the colour-related information content of a Red-Green-Blue (RGB) image, and which can be elaborated by means of proper multivariate analysis/feature selection techniques to easily identify those image features that are more useful to solve a specific problem. The reconstruction of the selected features as segmented images allows to evaluate in a critical manner the choices made automatically by the algorithm. In the present paper colourgrams are used for the detection of the red skin defect of raw hams, in order to render more objective and transferable the evaluation usually made by expert assessors. To this aim, after a preselection of 95 raw ham samples by a panel test, the corresponding RGB images were converted into colourgrams, which in turn were used to build classification models using Partial Least Squares-Discriminant Analysis (PLS-DA) and a Wavelet Packet Transform-based feature selection/classification algorithm (WPTER). Feature selection allowed to discriminate the defective samples using only three variables, with a Classification Efficiency in prediction of an external test set equal to 97.8%. The reconstruction of the samples images using only the selected features confirmed the reliability of the obtained classification model. Industrial Relevance: The evaluation of pig thighs is currently carried out by subjective methods, i.e. expert, long-trained personnel is needed to detect the presence or absence of defects. The method presented here would allow to uniform and drastically shorten the time needed for evaluation, and to avoid the main problems connected with human evaluation, i.e., subjectivity, possible unreliability, non-transferability and difficulty to collect historical data. Furthermore, it might represent a first step for setting up a comprehensive method of evaluation, aiming to take into account also other types of defects of raw hams destined to seasoning. More in general, thanks to its flexibility, this approach could be also successfully applied for the detection of other types of aspect-related features, even to monitor different kinds of products.