Error of the compensation fuzzy neural network fruit grader

This investigation analyses grading errors based on compensating fuzzy neural network to find out methods to improve citrus fruit grading. The first four harmonic components of discrete Fourier transform of citrus fruit images were sorted according to expected fuzzy values, and formed into four training sample sets which were in the order of monotone increasing, monotone decreasing, bell-shaped distribution and sawtooth distribution respectively. The new four sample sets were used to train compensating fuzzy neural network of the same architecture. Test set was sorted in monotone decreasing order. Test results show that model trained with monotone decreasing sample set has the smallest grading error which is 1.875%. Models trained with bell-shaped distribution, monotone increasing and sawtooth distribution have larger grading errors, which are 15%, 63.125%, 75% respectively. Grading error is modeled, and analyzed on correlation of grading error and differential order of sample sets. Results show that grading error is smaller with big correlation coefficient if training and testing sets are in the same order. On the contrary, if training and testing sample sets are in different orders, the correlation coefficient is smaller and the grading error is larger. This discloses testing and training sets shall be sorted in the same order according to qualitative feature as far as possible to improve the correlation coefficients. Performance of auto-grading system based on neural network model can be improved greatly by sorting samples in the same order.