Blueberry Bruise Detection by Pulse-Phase Thermography and Neural Network

Abstract.

During harvest and processing, blueberries suffer mechanical damage which reduces their market value. Mechanical damage which tears the skin or results in gross deformation can be detected by current vision-based sorting systems, but bruising may occur which is both visually undetectable and severe enough to reduce fruit quality. A rapid, non-destructive method for the detection of bruised blueberries is therefore needed. Pulse-phase thermography is a technique of infrared thermography which can produce data corresponding to the material properties of an examined object. In this study, bruised and unbruised blueberries of various treatments were analyzed by pulse-phase thermography and the resulting data were used to train multi-layer perceptrons. The trained models were then used to classify berries as bruised or unbruised, and their accuracy in this task was discussed. Under ideal conditions, multi-layer perceptrons classified berries with accuracies greater than 90%. For certain subsets, accuracy may be as low as 76%. Across the entire sample, accuracy was 83.5%. These results indicate that pulse-phase thermography could be a viable technique for the identification of bruised blueberries. Additionally, results suggest that this technique may be used to estimate the severity of bruising, but additional work in mitigating confounding variables is needed before this task can be conducted with appreciable accuracy.