Visual quality is an important factor for consumer purchasing decisions of fresh-cut lettuce. Consumer behavior towards produce quality has been studied via traditional human evaluations. For sensory studies, quality evaluations are commonly done by trained human panels and consumer panels. This study was to investigate the possibility of replacing human evaluation with a machine based approach, using image capturing and analysis, and determine whether efficiency of produce quality analysis can be enhanced. Three types of data were collected: (a) consumers’(n=200) evaluation of lettuce on the picture, (b) instrumental analysis of samples (package head-space gas composition (O2, CO2) and electrolyte leakage), (c) image analysis of lettuce on the pictures. For image analysis, ImagePro’s smart segment tool was used to classify the lettuce samples into five regions: adult leaf, baby leaf, rib, rib degradation, and leaf degradation. This was used to find and calculate L*a*b, hue angle, chroma values, area, and relative area of these regions. Lettuce samples consisted of four cultivars (Green Forest, King Henry, Parris Island Cos, PI 491224). Samples were measured on day 7, 10, and 13 of storage. To predict browning score, data sets (b) and (c) were fed into a regression algorithm. The scores assigned by trained panels served as the target variables. The results showed a strong correlation between consumer’s browning score on the pictures and predicted scores generated by the regression model (r=0.74). Interestingly, removing the instrumental data set (b) did not worsen the model’s performance. The model achieved an R2 of 0.92 and RASE of 8.90 when using trained data sets (a) and (c), and an R2 of 0.91 and RASE of 8.53 when using trained data set (c) only. While a correlation coefficient of 0.74 indicates a promising relationship between image analysis and human evaluation of browning score, it’s not sufficient to definitively replace human evaluation. Further studies with larger datasets and exploration of more advanced machine learning models could lead to a more robust statistical model.
