| 基于双分支注意力差分的扩散模型噪声抑制方法 |
投稿时间：2025-11-11 修订日期：2025-12-17  点此下载全文 |
| 引用本文： |
| 摘要点击次数: 17 |
| 全文下载次数: 0 |
|
| 基金项目:中国烟草总公司重点研发项目(110202202042)；基于数字孪生的数字化实验平台的研究及应用(AW2022027)。 |
|
| 中文摘要:多数的扩散模型使用Unet作为去噪过程的主干网络,一方面,注意力机制的应用显著提升了Unet的表现力,而另一方面,注意力机制在关注图像的主体信息之外,还会分配一部分权重给无效信息。这种对无效信息的关注限制了模型性能的进一步提升。基于以上问题,提出一种适用于Unet模型的双分支注意力差分方法,用于抑制Unet中注意力图对无效信息的关注。具体来说,在训练过程,模型的每一层都同时训练学习出两套注意力机制,用于得到两个具有类似特征的注意力权重矩阵,接着将两个注意力权重矩阵执行差分操作,并引入额外的超参数来控制注意力矩阵之间的差分强度,最后使用差分后的注意力权重得到最终的注意力图。我们在多个真实数据集中进行实验来证明双分支注意力差分机制的有效性,实验结果显示,融入该方法之后,多个主流模型的生成质量均能有效提升。 |
| 中文关键词:扩散模型 Unet 注意力机制 注意力差分 |
| |
| Diffusion Model Noise Suppression Method Based on Dual-Branch Attention Difference |
|
|
| Abstract:Most diffusion models utilize Unet as the backbone network for the denoising process. On one hand, the application of attention mechanisms has significantly enhanced the representational capacity of Unet; on the other hand, while focusing on the main content of the image, the attention mechanism also allocates a certain amount of weight to irrelevant information. This allocation to invalid information limits further improvement in model performance. To address this issue, we propose a dual-branch attention difference method suitable for Unet models, which suppresses the attention to irrelevant information in the attention maps. Specifically, during training, each layer of the model simultaneously learns two sets of attention mechanisms to produce two attention weight matrices with similar characteristics. These two matrices are then subjected to a difference operation, with an additional hyperparameter introduced to control the intensity of the difference between them. Finally, the differentiated attention weights are used to generate the final attention map. We conducted experiments across multiple real-world datasets to validate the effectiveness of the dual-branch differential attention mechanism. The results show that incorporating this method consistently enhances the generation quality of several mainstream models. |
| keywords:Diffusion Models Unet Attention Mechanism Differential Attention |
| 查看全文 查看/发表评论 下载pdf阅读器 |
|
|
|
次