A two-camera machine vision in predicting alpha-amylase activity in wheat

• A two-camera machine vision (MV) based on a neural network was developed.
• Color, texture, and shape and size of wheat kernels were used as the network inputs.
• Alpha-amylase activities in sound and sprout-damaged wheat were measured.
• At least 95% of the joined kernels were segmented automatically.
• The MV predicted alpha-amylase with an accuracy of 6913 U/L (rmse) and 0.72 (R2).

Sprout damage in wheat is a serious problem worldwide because damaged wheat kernels contain alpha-amylase, an enzyme that causes poor baking quality of wheat. A two-camera machine vision (MV) with a neural network was implemented to quantify alpha-amylase activity in wheat using 16 visual properties of the kernels. Kernels were separated at image level using the marker-controlled segmentation algorithm before the properties (color, texture, and shape and size) of dorsal and ventral sides of kernels were extracted. Alpha-amylase activity in wheat was assessed analytically. The neural networks were trained, validated, and tested using the visual properties as the inputs and alpha-amylase activity as the output. The trained neural network predicted alpha-amylase activity with an accuracy of 6913 U/L (rmse) and R2 value of 0.72 for the wheat samples with alpha-amylase activity ranging over 178 to 28935 (U/L). Differences between visual properties of wheat samples calculated from the top and the bottom images was less than 0.5%. Light stability in time and influence of temperature on the cameras' color stability were less than 2% of the mean values. The challenges associated with the system, and recommendations to improve the system accuracy and robustness, and to decrease the system cost are presented.