NEURAL-NETWORK PREDICTION MODELING BASED ON ELASTOGRAPHIC TEXTURAL FEATURES FOR MEAT QUALITY EVALUATION

Multilayer feedforward neural networks are proposed to capture the non linearity between the system inputs and outputs to predict meat qualit y with the wavelet textural features fron the ultrasonic elastograms. This article investigated the efficiency of the training processes and the generalization of the networks using the gradient descent and Lev enberg-Marquardt algorithms in backpropagation. It was Sound that for this application in the case of low epoch training (below several thou sand epochs) using the gradient descent algorithm, the Levenberg-Marqu ardt algorithm was less efficient and in the case of high epoch traini ng (above several thousand epochs) using the gradient descent algorith m, the Levenberg-Marquardt algorithm was more efficient In the case of difficult convergence in the gradient descent algorithm the Levenberg -Marquardt algorithm converged effectively. In either case, the Levenb erg-Marquardt algorithm better modeled output variation accounting and network generalization. Weight-decay was further used in the Levenber g-Marquardt backpropagation to improve the generalization of the netwo rk models. The leave-one-out procedure was built into every training p rocess to ensure sufficient modeling on limited samples.