Optimization of U-Net Network and Attention Mechanism in Football Robot Image Segmentation

This paper addresses the semantic segmentation issues of the traditional U-Net in complex scenes, systematically studying the impact of various backbone networks and optimizers on segmentation accuracy. Experiments utilized ResNet50, MobileNetV2, and EfficientNet-B0 as backbone networks, comparing their loss and accuracy during training and validation. The results indicate that ResNet50 performed best in validation accuracy. Additionally, a comparison of the Adam, AdamW, and NAdam optimizers revealed that Adam had the fastest convergence rate, NAdam achieved the highest validation accuracy, and AdamW resulted in the lowest validation loss. Furthermore, by incorporating the Criss-Cross Attention (CCA) module in the bottleneck layer, the model's validation accuracy was enhanced to 0.9680. Ultimately, the model based on ResNet50-U-Net incorporating the CCA attention module and Adam optimizer achieved high-precision segmentation results on the test set.