ALLNet:面向退化图像的多任务密集预测
ALLNet: Multi-task Dense Prediction for Degraded Images
把 MoE 动态路由用于退化特征增强,并与任务特定特征做双向协同精炼,为在恶劣成像条件下做多任务密集预测(深度、语义分割等)提供了可复用的「复原-任务联合」框架。
原文依据 we designa MoE-basedMixtureofAdaptiveExperts(MaE)restorationcomponentnetworkthatenhancesdegradationfeaturesthroughdynamicrouting
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论多任务密集预测旨在通过统一网络同时处理多个像素级任务,以实现视觉场景理解。然而,恶劣的环境条件限制了此类任务的泛化性与实用性。为此,我们提出 ALLNet,一种有效挖掘退化模式并整合多任务协同信息的新框架。具体而言,我们设计了一个基于混合专家(MoE)的自适应专家混合(MaE)复原组件网络,通过动态路由增强退化特征,并引导任务特定特征的提取。此外,我们构建了任务感知协同精炼(TCR)模块,以捕获全局语义关联与跨任务依赖关系,促进退化图像上复原特征与任务特定特征之间的双向协同。据我们所知,这是图像退化条件下多任务密集预测的首次尝试。在退化版 NYUDv2 与 PASCAL-Context 基准上的实验结果表明,我们的架构在退化场景中显著优于现有方法。
英文原文摘要 1044 字符
Multi-taskdensepredictionaims tosimultaneously addressmultiplepixel-level tasksthroughaunifiednetworkforvisualsceneunderstanding.However,adverse environmental conditions limit thegeneralizationand practicality of such tasks. Toaddress this, we proposeALLNet, anovel framework that effectively explores degradationpatterns and integratesmulti-task collaborative information. Specifically, we designa MoE-basedMixtureofAdaptiveExperts(MaE)restorationcomponentnetworkthatenhancesdegradationfeaturesthroughdynamicroutingandguidestask-specific featureextraction. Furthermore,weformulateaTaskawareCollaborativeRefinement (TCR)moduletocaptureglobalsemanticcorrelationsandcross-taskdependencies,facilitatingbidirectionalcollaborationbetween restorationandtask-specific featuresondegradedimages. Tothebestofourknowledge, thisrepresentsthe firstattemptatmulti-taskdensepredictionunderimage degradation.ExperimentalresultsondegradedNYUDv2andPASCAL-Contextbenchmarksdemonstratethat ourarchitecturesignificantlyoutperformsexistingmethodsindegradedscenarios.