Based on seedling properties and stage of growth for cucurbitaceous and solanaceous vegetables, separate robots are being marketed for each. Full automatic grafting robots are used for solanaceous vegetables like tomato and egg-plant employing ordinary splice method by making a diagonal cut through the hypocotyl of both the scion and the rootstock. However, cutting one piece of cotyledon diagonally from the rootstock does grafting of cucurbitaceous vegetables like cucumber, melon, and pumpkin. This method had the advantage of easy recovery and high survival rate of seedlings. Only semi-automatic robots are marketed for this kind of plants because a fixed cotyledon orientation is required for grafting operation. Both the scion and the rootstock are loaded manually to their corresponding feeding devices. To replace the manual loading operation, this study proposed a neural network based automatic seedling loading system. The system automatically estimates the quality and determines the cotyledon orientation of seedling for guiding the loading device of the grafting robot. As a first step toward solution, we report the development of a model for seedling quality estimation and orientation detection using image processing and neural network techniques. The model has a learning ability and can judge seedlings according to the training patterns. A seedling leaves feature extraction model of 10 characteristics was proposed and a three-layer neural network was constructed. The experimental results indicate that the seedling leaves orientation was accurately detected with an average error of 3 degrees within 360 degrees of freedom and the machine vision system could properly classify seedlings into three classes (A-good, B-fair, and C-bad) according to the training pattern.