.png){kind=icon}
Return
Superpixel-informed transformer for multi-label image classification
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2026.133995.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Multi-label image classification aims to assign a set of relevant semantic labels to an image. However, current methods often suffer from region evidence confusion, where cues from small objects are diluted by dominant regions. In addition to this visual ambiguity, decoupled modeling remains a common hurdle, as it separates visual feature extraction from label relationship learning. To address these limitations, we propose the Superpixel-informed Transformer (SPiT), a novel end-to-end framework. SPiT adaptively decomposes an image into semantically coherent quasi-object regions using a differentiable superpixel network. Upon these regions, a class-query-based Transformer then performs unified evidence aggregation and label dependency modeling. Experimental results on PASCAL VOC 2007, MS-COCO 2014, and ISIC 2018 demonstrate the superiority of the proposed method, achieving mAP scores of 93.06%, 81.73%, and 56.72%, respectively, outperforming state-of-the-art techniques on all three benchmarks.
Keywords:
multi-label image classification
superpixel
Transformer
visual ambiguity
label dependency modeling
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W
Organization
Cited Papers
.png){kind=icon}
ShareSave
ShareSaveShareSaveShareSave
View more