Comparison of colour and texture feature extraction methods to predict anthocyanins content in Sangiovese grapes

Colour and texture are the two main sources of information contained in RGB images of food products. Different image-level approaches are available to analyse the image properties based on the extraction of colour and texture features, and the selection of the most appropriate method is a critical point, since it could significantly impact the outcomes. The present study has three main objectives. Firstly, we propose an innovative data dimensionality reduction method to extract and codify the texture features of an RGB image into a one-dimensional signal, named texturegram (TXG). Then, TXG approach is compared with different image-level feature extraction methods, such as colourgrams (CLG), Soft Colour Texture Descriptors (SCTD) and Grey Level Co-occurrence Matrices (GLCM). These techniques were used to analyse a benchmark dataset of RGB images already considered in a previous study to build Partial Least Squares (PLS) models and relate the image features with anthocyanins content of red grape samples. We also investigated the possible advantages of combining the colour and texture information brought by the different image-level techniques using data fusion. PLS models were calculated considering different partitions of the RGB image dataset into training and test set. The performances of the different models were statistically evaluated by means of Analysis of Variance (ANOVA) and Principal Component Analysis (PCA). Overall, the results suggested an interesting, even if slight, improvement of the model performances when fusing CLG and TXG, but also highlighted the hybrid nature of TXG to simultaneously explore colour and texture properties.