The use of trained artificial neural networks (ANNs) for agricultural processing, handling, and process control, such as pattern recognition, classification and weight prediction, offers potential for multi-dimensional function fittings and enhanced accuracy in machine-vision based procedures. In this study, optimization of ANNs for machine vision based applications for better prediction accuracy has been conducted using soybean weighing as an example. Neural network systems consisting of a varying number of neurons trained under dissimilar algorithms were compared in determining the weights of soybeans based on the correlation of weight to features extracted from one- and two-direction images. Results show that imaging from the side of a soybean produces superior data to that of top-down images, and that with a properly trained neural network, weight predictions could be accurate up to a relative error of less than three percent. Furthermore, the continuous dependence of weight to features of the soybean suggested use of a training batch consisting of uniformly distributed weights.