Laser scattering imaging combined with CNNs to model the textural variability in a vegetable food tissue

In this work, the capacity of the laser scattering imaging, combined with predesigned convolutional neural networks was studied to model the textural variability of pre-processed banana as vegetable tissue. Three different imaging embedders (Inception V3, Painters and VGG19) and four types of images (based on colour channels RGB) were tested for that purpose. Two batches of bananas were stored for 0, 3, 7 and 10 days. Fruits were sliced and analysed in imaging and firmness terms. Texture differences were observed for different storage times and fruit zone. Firmness mean and dispersion decreased with storage time. Images from samples revealed light-transmission decreased with storage time. After multivariate statistical procedures, imaging data showed similar behaviour to firmness concerning dispersion. Regression studies evidenced dependency between imaging and firmness data, albeit different errors were obtained depending on the convolutional network and image types. Moreover, after dividing samples into four different models of multilevel categories, successful predictions were obtained from all networks in function of the used image type. Overall, information from the R channel improved results for statistical analytics. However, Inception V3 and Painters also present good results from isolated information of G and B. Results evidenced the capacity of the used imaging procedure to capture and model the variance generated in texture due to tissue differences produced by both storage process and fruit zone.

Automated summary

This study investigated the capacity of laser scattering imaging combined with predesigned convolutional neural networks to model the textural variability of pre-processed banana as vegetable tissue. The results showed a correlation between imaging data and firmness data, with different errors depending on the convolutional network and image types. Additionally, successful predictions were obtained using different image types in multilevel models.