Transfer Learning and Data Augmentation Analysis for Computer Vision in Robot Football using Deep Learning

Abstract

Robot Soccer is a promising field for advances in computing and robotics, especially in computer vision with Convolutional Neural Networks (CNN). New models in this area can optimize the intelligence and autonomy of machine learning systems. This study developed databases to create Image Classification and Object Detection models in humanoid robot football, analyzing the impact of Transfer Learning (TL) and Data Augmentation (DA). Using the Edge Impulse platform, experiments with 5 image classes were carried out for Image Classification, while 4 databases with different image sizes were used for Object Detection. The results showed that TL increased the accuracy of the Classification models between 11.25% and 38%, and the combination with DA resulted in variations of up to 3.5% in accuracy in comparison to when applying only TL. In Object Detection, increasing the number of images improved accuracy by 33% and Recall by 37%. These findings highlight the importance of these techniques for the learning and interaction of robotic agents in complex environments. Future studies can focus on tuning hyperparameters and new architectures to optimize these models.