Fields such as early education and motor rehabilitation provide a space for their integration with augmented reality devices as AR-Sandbox, in order to provide support to these fields, generating a feedback of tasks carried out through the recognition of images based on convolutional neural networks. However, the nature of the AR-Sandbox generates a high noise level of the acquired images, for this reason the present study has as purpose the implementation and comparison of three segmentation methods (Canny Edge Detector, Color-space and Threshold) for the training and prediction phase of a convolutional neural network model previously established. When carrying out this study, it was obtained that the combined model with color-space segmentation presents an average percentage of 99% performance for the classification of vowels, described by the AUC of the ROC curve, this being the model with the best performance.