在日常生活中,口罩遮挡口部和鼻部、眼镜遮挡眼睛等情况对基于人脸的年龄估计模型造成了较大的挑战。为应对这一问题,本文设计了一种基于知识蒸馏的自监督特征重建模型。由于缺少包含年龄标记的遮挡人脸数据集,我们选择在无遮挡的原始人脸图像数据集上合成口罩和眼镜作为遮挡数据集,并通过自监督学习的方式实现特征重建。在方法中,我们使用主流的年龄估计模型作为教师模型,训练一个结构相同但附加了特征重建模块的学生模型。通过知识蒸馏,将教师模型的知识迁移至学生模型,并利用特征重建模块对遮挡区域进行特征补全,以提升模型对遮挡人脸图像的年龄估计效果。实验结果表明,提出的方法显著提升了模型在遮挡图像上的年龄估计性能。特征重建模块有效缓解了常见遮挡对年龄估计的负面影响,从而增强了模型的实用性和鲁棒性。In daily life, situations where masks cover the mouth and nose, and glasses cover the eyes, pose significant challenges to facial based age estimation models. To improve the robustness of the facial age estimation model under occlusion, this paper proposes a self-supervised feature reconstruction model based on knowledge distillation. Due to the lack of age labeled occluded face datasets, we chose to synthesize masks and glasses as occluded datasets on the original unobstructed face image dataset, and achieved feature reconstruction through self-supervised learning. In the method, we use mainstream age estimation models as teacher models and train a student model with the same structure but additional feature reconstruction modules. By knowledge distillation, the knowledge of the teacher model is transferred to the student model, and the feature reconstruction module is used to complete the features of the occluded area, in order to improve the age estimation effect of the model on occluded face images. The experimental results show that the proposed method significantly improves the age estimation performance of the model on occluded images. The feature reconstruction module effectively alleviates the negative impact of common occlusion on age estimation, thereby enhancing the practicality and robustness of the model.
在乳腺癌的诊断与治疗过程中,乳腺肿瘤影像分割技术扮演着至关重要的角色,其精确度直接关系到病理分析的准确性及临床决策的有效性。近年来,U-Net及其改进模型在乳腺影像分割领域取得了显著的进展。U-Net的编码–解码结构和跳跃连接设计在提取多尺度特征和保持分辨率方面展现出独特优势,已发展成为医学图像分割领域的经典方法。随着研究的不断深入,针对U型网络的多方面优化进一步提升了其在乳腺医学图像分割中的性能。此外,U-Net在多模态影像分割任务中的应用也逐渐扩展。本文综述了基于U-Net的乳腺肿瘤分割模型的研究现状,探讨了其在数据集构建、性能评估指标、网络结构优化以及实际应用中的最新进展,并分析了当前研究面临的挑战和未来发展方向。该综述旨在为乳腺肿瘤影像分割领域的研究和应用提供重要的参考。Breast tumor image segmentation is a pivotal technology in the diagnosis and treatment of breast cancer, with segmentation accuracy directly influencing subsequent pathological analysis and clinical decision-making. In recent years, U-Net and its improved models have achieved significant advancements in the field of breast image segmentation. The encoding-decoding structure and skip connection design of U-Net offer unique advantages in extracting multi-scale features and maintaining resolution, establishing it as a classic method for medical image segmentation. As research progresses, various optimizations of the U-Net network have further enhanced its performance in breast medical image segmentation. Moreover, the application of U-Net in multi-modal image segmentation tasks has also gradually expanded. This paper provides a comprehensive review of the research status of U-Net-based breast tumor segmentation models, discussing the latest advancements in datasets, performance metrics, network structure improvements, and practical applications, while also analyzing current research challenges and future development directions. This review serves as an important reference for the research and application of breast tumor image segmentation.
时间序列相似检测在金融数据、电力数据挖掘等场景都有很重要的作用。为了解决时间序列深度哈希网络存在哈希量化损失的问题,提出一种端到端的深度对比学习时间序列哈希网络(Deep Contrastive Time Series Hash,DCTSH)。通过引入自适应...
详细信息
时间序列相似检测在金融数据、电力数据挖掘等场景都有很重要的作用。为了解决时间序列深度哈希网络存在哈希量化损失的问题,提出一种端到端的深度对比学习时间序列哈希网络(Deep Contrastive Time Series Hash,DCTSH)。通过引入自适应二值化网络与哈希损失,消除二值化哈希时的量化误差,使得模型端到端训练生成的时间序列哈希编码,具有更好的表达效果与泛化能力。针对无标签时间序列数据,通过聚类改进对比学习网络的负样本选择来增强时间序列表示学习能力。在多个时间序列数据集上实验结果表明,DCTSH相较于之前的方法检测精度显著提升。
暂无评论