The performance of semi-supervised semantic segmentation models is significantly influenced by pseudo-label. To enhance the quality of pseudo-labels, we propose the Hybrid Architectures Ensemble Learning (HAEL) method. Specifically, we observe that different network architectures excel in specific tasks. CNN-based models are adept at capturing fine-grained features, whereas ViT-based models excel in perceiving significant intra-class variations and long-range dependencies. To integrate the diversity and specificity of these architectures, we propose a Hybrid Teacher–Student Model (HTM) incorporating two teacher–student branches. Each branch dynamically adjusts its prediction weights based on its expertise, thereby refining pseudo-labels by enhancing fine-grained feature representations and long-range dependencies. Furthermore, to enhance pseudo-labels’ stability and encourage them to learn model-specific features, we introduce a Similarity-Guided Channel Dropout (SCD) mechanism. This mechanism retains features shared by teachers with higher probability, mitigating the degradation of pseudo-label quality associated with the loss of important feature channels discarded by classical channel dropout. Meanwhile, unlike HTM, which explicitly combines the masks of the two architectures based on confidence maps, SCD implicitly explores model-specific features through rigorous screening based on each teacher’s unique knowledge. Finally, a Blind Spot Constraint (BSC) is introduced to impose additional penalties to improve the accuracy of hard samples in pseudo-labels. Extensive experiments on the Pascal VOC 2012 and Cityscapes demonstrate the effectiveness of the HAEL, achieving significant performance improvements.

中文翻译：

---

混合架构集成学习，用于半监督分割中的伪标签细化


半监督语义分割模型的性能受伪标签的显著影响。为了提高伪标签的质量，我们提出了混合架构集成学习 （HAEL） 方法。具体来说，我们观察到不同的网络架构在特定任务中表现出色。基于 CNN 的模型擅长捕获细粒度特征，而基于 ViT 的模型擅长感知显著的类内变化和长期依赖关系。为了整合这些架构的多样性和特异性，我们提出了一种混合师生模型 （HTM），其中包含两个师生分支。每个分支都会根据其专业知识动态调整其预测权重，从而通过增强细粒度特征表示和长期依赖关系来优化伪标签。此外，为了提高伪标签的稳定性并鼓励它们学习特定于模型的特征，我们引入了一种相似性引导通道丢失 （SCD） 机制。这种机制以更高的概率保留了教师共享的特征，减轻了与经典通道丢失丢弃的重要特征通道相关的伪标签质量的下降。同时，与 HTM 基于置信度图显式组合两种架构的掩码不同，SCD 根据每位教师的独特知识，通过严格的筛选隐式探索模型特定的特征。最后，引入了盲点约束 （BSC） 来施加额外的惩罚，以提高伪标签中硬样本的准确性。在 Pascal VOC 2012 和 Cityscapes 上进行的广泛实验证明了 HAEL 的有效性，实现了显著的性能改进。