A review on image semantic segmentation based on domain adaptation
With the rapid development of deep learning,semantic segmentation algorithms have seensignificant performance improvements. However,they heavily rely on large-scale paired image datas-ets and labor-intensive pixel-level annotations. Artificially synthesized images,characterized by their scalability and ease of annotation,effectively reduce training costs by replacing real images for train-ing. Nonetheless,the domain gap between synthetic and real images influences the generalization capa-bility of segmentation networks. To address this issue,Domain Adaptive Semantic Segmentation (DASS) algorithms aim to extract domain-invariant features,thereby minimizing domain gaps and en-hancing the network generalization on the target domain. This paper classifies mainstream DASS algo-rithms according to the network structure,analyzes the performance comparison results of different al-gorithms,and proposes future research directions. The results show that early DASS methods utilize generative adversarial networks to align the distribution between the source and target domains. How-ever,their network structure is complex,achieving only global alignment and unable to realize fine alignment between different categories,resulting in lower performance. Subsequent methods gradually turn to the self-training networks,utilizing pre-trained segmentation network to generate pseudo la-bels in the target domain,providing supervision for the next round of training. This approach has a sim-pler structure and better performance. With the advent of Transformer,their powerful feature extrac-tion capability further improves the accuracy of existing DASS methods.
万方数据
维普
