AIxiv专栏是机器之心发布学术、技术内容的栏目。过去数年,机器之心AIxiv专栏接收报道了2000多篇内容,覆盖全球各大高校与企业的顶级实验室,有效促进了学术交流与传播。如果您有优秀的工作想要分享,欢迎投稿或者联系报道。投稿邮箱:liyazhou@jiqizhixin.com;zhaoyunfeng@jiqizhixin.com。
论文链接:https://arxiv.org/abs/2403.12494 代码链接:https://github.com/YangSun22/TC-MoA 论文题目:Task-Customized Mixture of Adapters for General Image Fusion
我们提出了一个统一的通用图像融合模型,提供了一种新的任务定制混合适配器(TC-MoA)用于自适应多源图像融合(受益于动态聚合各自模式的有效信息)。 我们为适配器提出了一种互信息正则化方法,这使得我们的模型能够更准确地识别不同源图像的主导强度。 据我们所知,我们首次提出了一种基于 MoE 的灵活适配器。通过只添加 2.8% 的可学习参数,我们的模型可以处理许多融合任务。大量的实验证明了我们的竞争方法的优势,同时显示了显著的可控性和泛化性。
![图片](https://image.jiqizhixin.com/uploads/editor/afb60fee-6246-405f-a206-5e0c77ebd175/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/ca843c44-7a6e-4e7b-bb69-bd53ead2af39/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/eda3952e-96ea-4a02-8912-a85514d9fdc1/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/36456dc8-b136-4e73-966b-5bbe4be0ae2b/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/96b8df24-e7e7-4980-b1ec-89b112127768/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/b78e312f-bf2c-4461-80a4-3208169641ed/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/e05b19c6-3736-4ad9-9667-45a6f93a02c1/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/c4f21aa3-5927-4208-99b8-253bcbbb0340/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/1948d074-6dbf-447f-a3cf-afc5ac1b8420/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/a65c117b-b1e0-459c-80ec-1da0c416d764/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/047c42cd-b568-482d-8a24-385b99d415d2/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/8a461f89-7298-4412-ad89-8675a7b48b27/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/1f81e73e-53d0-4a62-9162-7124b3ac0b31/640.png)