Estimation of weight percentage of scabby wheat kernels using an automatic machine vision and neural network based system

In examining the quality of wheat for pricing purposes, or the scab-resistance of wheat as required for breeding research, it is important to rapidly determine the weight percentage of scabby kernels (WPSK). This study involved development of an automatic system to perform such tasks. This system is based on the color features of scabby kernels captured by a machine vision system. The color features were processed to produce numerical values that were correlated to WPSK using a neural network. Schemes were developed to synchronize an automatic sample feeder with image capturing, followed by automatic image processing and neural network computing. Wheat kernels were distributed in a single layer for image capturing, which minimized the random errors caused by overlapping of kernels and eliminated the need to acquire multiple images to representa single sample when kernels are distributed in multiple overlapping layers. Statistical analysis indicates that the correlation coefficient between estimated WPSK and actual WPSK was 0.96, with a mean absolute error of 1.32% and maximum absolute error of 5.22%. The system could be stabilized through an online color-compensation procedure that dealt with illumination variations.