Semantic segmentation for soccer robots based on the U-Net model

As one of the largest and most comprehensive robotics competitions both domestically and internationally, the RoboCup competition has attracted an increasing number of university students to participate. The core of vision systems for soccer robots lies in semantic segmentation. In recent years, with the rapid development of deep learning techniques, image segmentation methods based on Convolutional Neural Networks (CNNs) have made significant progress. The U-Net model, featuring a contracting path, bottleneck, and expansive path, has been used to achieve precise segmentation and has demonstrated good performance in image segmentation tasks. However, for images with complex backgrounds and diverse targets, traditional U-Net models often struggle to achieve ideal segmentation results. This paper proposes an image segmentation model based on the ResNet101 and U-Net architecture to address the visual semantic challenges faced by soccer robots. Through a comparison of multiple models and optimizers, ResNet101 is ultimately selected as the feature extractor, and Nadam is chosen as the optimizer. This approach combines the encoder-decoder structure of U-Net with skip connections to capture multi-scale feature information in images. To further enhance the model's performance, data augmentation is incorporated, and an attention mechanism is introduced into specific layers of ResNet101, achieving significant improvements in target segmentation accuracy and generalization capability.