【摘要】 针对中医舌象形态复杂、特征多样、标注困难等问题，提出一种基于半监督多任务协同的舌象特征智能分类方法（GlossoSynthNet）。方法引入双向对称编解码结构与最小熵约束的半监督学习机制实现舌体注意力区域的自动提取；结合空洞空间金字塔池化模块与伪标签驱动策略完成舌苔与舌体的分离；构建融合多尺度特征的多任务分类结构实现舌象特征的多指标联合识别。通过多中心中医医院采集3123例患者样本构建舌象图像数据集，实验结果表明，所提出的GlossoSynthNet在分类性能方面显著优于现有主流舌象特征分类方法。更多还原

【Abstract】 To address the challenges of complex morphology, diverse features, and annotation scarcity in traditional Chinese medicine(TCM) tongue image analysis, an intelligent tongue feature classification method based on semi-supervised multi-task collaboration, named GlossoSynthNet, was proposed. A bidirectional symmetric encoder-decoder structure was introduced, coupled with a semi-supervised learning mechanism constrained by minimum entropy, to achieve automatic extraction of tongue attention regions. An atrous spatial pyramid pooling(ASPP) module and a pseudo-label-driven mechanism were incorporated to achieve the separation of tongue coating and body. A multi-task classification framework integrating multi-scale features was constructed to enable joint recognition of multiple tongue diagnostic indicators. A dataset consisting of 3123 tongue images was collected from multiple TCM hospitals. Experimental results demonstrate that the proposed GlossoSynthNet significantly outperforms existing mainstream methods in terms of classification performance.更多还原