基于跨层级注意力学习的RGB-T显著目标检测

打开文本图片集

DOI：10.13705/j. issn.1671-6841.2023163

RGB-T Salient Object Detection Based on Cross-level Attention Learning

WEI Mingjun 1，2 ，WEI Shuai'，LIU Yazhi 1，2 ， LI Hui1 （1.College of Artificial Intelligence，North China University of Science and Technology， Tangshan 063210， China；2. Hebei Provincial Key Laboratory of Industrial Inteligent Perception， Tangshan O63210，China）

Abstract： RGB-thermal saliency object detection （RGB-T SOD） aimed to segment common salient regions in both visible light images and corresponding thermal infrared images.To address the problem of insuffcient utilization of cross-level complementary information among existing methods，a cross-level feature attention learning network （CALNet）was proposed for the RGB-T SOD task. Specifically，the network included a cross-level attention learning module（CAL），which used non-local attention to interact cross-level information among multiple modalities and could fuly explore global positions and local details across diffrent modalities and levels.Inaddition，the network also introduced a global information module（GIM）and a multi-interaction module（MIB），both of which could model and explore multi-type information in a layer-by-layer decoding process for more accurate RGB-T SOD. Extensive experiments on public RGB-T datasets demonstrated that the proposed network achieved excellent performance compared with state-of-the-art algorithms in the field.

Key words： multimodal; non-local attention; RGB-T; salient object detection; feature fusion

0 引言

显著目标检测（salient object detection，SOD）旨在从视觉场景中准确地检测和分割最具吸引力的物体，在各种计算机视觉研究中，如视频对象分割、光场图像分割、目标跟踪和实例分割²等，都起着重要作用。（剩余9317字）