Optimisation of Semantic Segmentation based on Unsupervised Domain Adaptation

Recently, techniques such as deep learning and neural networks have been well performed and widely used in computer vision tasks such as image recognition, object detection, and semantic segmentation. However, these models often require a lot of labeled data, while also ensuring that the source and target domains are broadly similar in feature distribution, which is often not satisfied in real-world applications. This paper explores the potential of Unsupervised Domain Adaptation (UDA) methods in addressing these challenges, especially in semantic segmentation tasks. We also improve the quality and training stability of the generated images through a CycleGAN-based image transformation method that improves the model network architecture and optimises the loss function. Then, we explore the effect of domain adaptation on semantic segmentation by comparing the semantic segmentation results of the source-only model and domain-adapted model on the same training dataset. It is found that the unsupervised domain adaptation model improves the overall performance compared to the source model, with a slight increase in mIoU value, achieves significant performance improvement in several categories, and can effectively narrow the gap between domains, which can be combined with more advanced feature alignment strategies to better enhance model generalization and its performance in real-world scenarios.