Hierarchical Hybrid Transformer for Image Super-Resolution

Basaran B., Polat D., Yilmaz G. N., Berkol A.

8th International Congress on Human-Computer Interaction, Optimization and Robotic Applications, ICHORA 2026, Ankara, Turkey, 21 - 23 May 2026, (Full Text)

  • Publication Type: Conference Paper / Full Text
  • Doi Number: 10.1109/ichora69329.2026.11537261
  • City: Ankara
  • Country: Turkey
  • Keywords: deep learning, hierarchical feature extraction, hybrid neural networks, image quality enhancement, image restoration, swin transformer, vision transformers
  • TED University Affiliated: Yes

Abstract

This study proposes five lightweight image SuperResolution (SR) architectures, Restormer, SwinIR, HINet, and two hybrid ensemble models (Hybrid-1 and Hybrid-2), to restore heavily compressed JPEG images. The proposed Hybrid-2 model achieves state-of-the-art performance on the LIVE1 dataset with only 41,347 parameters, attaining Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) values of 37.68 dB and 0.9921, outperforming existing approaches including Swin2SR, ART, and ART+. These results demonstrate that strategically combining complementary lightweight architectures can surpass models with millions of parameters on highly degraded images, enabling practical deployment on mobile, embedded, and Internet of Things (IoT) platforms.