Kong Kuok YongTan Fong Ang0000-0003-0793-3308Chin Soon KuFirdaus SahranLip Yee Por2026-10-162026-10-16202510.1109/ACCESS.2025.3602481https://dspace-cris.utar.edu.my/handle/123456789/11765Zero-shot semantic segmentation continues to face challenges in effectively handling unseen object classes, despite its critical applications in medical imaging, autonomous driving, and aerial imagery. This paper introduces a hybrid approach that combines three key components: a Swin Transformer backbone, a Flash Masked Attention Transformer Decoder, and a Deformable Pixel Decoder, to enhance zero-shot semantic segmentation accuracy. The Swin Transformer backbone extracts hierarchical image features, the Flash Masked Attention Transformer Decoder refines feature extraction by focusing cross-attention on foreground regions, and the Deformable Pixel Decoder specializes in capturing small object features. Extensive evaluations on benchmark datasets demonstrate the method’s performance, achieving scores of 31.1 mIOU and 71.8 pAcc on ADE20K, 9.3 mIOU and 54.4 pAcc on ADE20K-Full, and 94.7 mIOU and 97.5 pAcc on Pascal VOC, significantly outperforming existing approaches. These results establish the effectiveness of the proposed method in advancing zero-shot semantic segmentation, particularly in scenarios involving diverse and previously unseen object classes. © 2013 IEEE.en-USclipSemantic segmentationswim transformervisual transformerzero-shotBenchmarkingDecodingMedical image processingPixelsSemantic SegmentationSemanticsAerial imageryAutonomous drivingClipCritical applicationsHybrid approachObject classSwim transformerVisual transformerZero-shotAntennasjournal-article