Optimizing semantic segmentation tasks for soccer robots: an empirical study on deep learning model selection and hyperparameter tunings

This study aims to enhance the performance and applicability of soccer robots in semantic segmentation tasks. Initially, we prepared the data and selected three models, DenseNet121, ResNet50, and Mobilenetv2, for training. During the training process, we employed data augmentation methods such as image rotation, horizontal flipping, and cropping, and experimented with various loss functions and optimizers. The experimental results showed that the combination of Mean Squared Error (MSE) as the loss function and Adam as the optimizer performed best. Additionally, we explored the addition of attention mechanisms and freezing layers, and introduced a dynamic learning rate strategy. In terms of model selection, MobileNetv2 was considered the optimal model due to its high validation accuracy, reasonable training time, and resource usage. In the final stage, we trained the final model based on the excellent parameters and models summarized earlier, achieving a validation accuracy of 94.31% in the last epoch. Overall, our research provides an effective strategy to optimize the semantic segmentation tasks of soccer robots.