Immersive techniques such as augmented reality through devices such as the AR-Sandbox and deep learning through convolutional neural networks (CNN) provide an environment that is potentially applicable for motor rehabilitation and early education. However, given the orientation towards the creation of topographic models and the form of representation of the AR-Sandbox, the classification of images is complicated by the amount of noise that is generated in each capture. For this reason, this research has the purpose of establishing a model of a CNN for the classification of geometric figures by optimizing hyperparameters using Random Search, evaluating the impact of the implementation of a previous phase of color-space segmentation to a set of tests captured from the AR-Sandbox, and evaluating this type of segmentation using similarity indexes such as Jaccard and Sørensen-Dice. The aim of the proposed scheme is to improve the identification and extraction of characteristics of the geometric figures. Using the proposed method, an average decrease of 39.45% to a function of loss and an increase of 14.83% on average in the percentage of correct answers is presented, concluding that the selected CNN model increased its performance by applying color-space segmentation in a phase that was prior to the prediction, given the nature of multiple pigmentation of the AR-Sandbox.