System Design for Plant Factory Operation Based on Three-dimensional Data Base (4): Recognition of 3-dimentional Image by Using Neural Network

When we think about automation in a plant growth factory, development of intelligent robots is expected. The vision of intelligent robots and the recognition using this sight require the study of a recognition system based on artificial intelligence. In this project we examined a recognition system using 3-Dimensional range image data processed by a neural network. Especially, we investigated how to make good input for a neural network from 3-D data. This input should contain maximal information with as small input-data as possible. In this study, we examined orange and eggplant. Input to the neural network were 64 vector lengths from the center of the fruits to the surface. Neural network was trained using the back propagation rule. Hidden layer neuron number and output neuron number were 2. Desired was output 1-0 for the eggplant and 0-1 for the orange. The number of training examples was 25 both in orange and eggplant. Number of interations needed to train the neural network was 10, 000. Verification of the performance of the recognition showed very well results. The recognition system was able to distinguish an orange from an eggplant with great certainty. This system was intended to recognize round objects. Improvement of algorithm for recognition of long and curved objects like cucumber and uneven surfaces like pimento is the next step. Further, based on the vector lengths, we are developing a recognition system for cracks at the fruit surface.