Abstract:

A machine vision based method was developed as an alternative to the physical and chemical methods and the
manual method to detect the fermentation process of tempeh. The texture characteristics of the tempeh images taken during
fermentation were extracted by calculating the gray level co-occurrence matrix (GLCM) for HSI (hue, saturation, intensity)
color model. Compared with the manual sensory evaluation method, the turning point of texture feature curves of tempeh
images could be better used as decision basis to distinguish four fermentation periods and further divide them into six stages.
Texture data analysis showed that the starting point of tempeh fermentation by Rhizopus oligosporus was determined to be
3 h earlier than by using the manual method, and the extreme point of texture feature curves of tempeh images represented
the end point of tempeh fermentation. Texture feature time series was developed by sliding the observation window, and
then a nonlinear time series model for tempeh fermentation process was established by neural networks ensemble training.
Finally, the extreme points of texture feature curves were predicted using the model, achieving the analysis of the solid-state
fermentation process of tempeh and the prediction of its end point.

Key words: tempeh, solid-state fermentation, time series, neural network ensemble