A Context Aware Conformer for Single Image Super-Resolution: Merging Global Attention With Incremental Features

Single Image Super-Resolution (SISR) plays a key role in next-generation consumer electronics like smartphones, smart TVs, and VR devices, where clear and detailed images are essential. Transformer-based methods effectively extract global features but often struggle with fine detail extraction and adapting to changing contexts. To address this, we propose a Context-Aware Conformer Network (CACN) for SISR. The CACN combines Convolutional Neural Networks (CNNs) and transformer blocks, called Adaptive Context Transformer Blocks (ACTBs), due to their context-aware adaptation capabilities. These ACTBs are cascaded to form lightweight or deep CACN variants. Each ACTB consists of Cascaded Context Aware Transformers (CCATs) with an Adaptive Channel Attention and Spatial Attention Fusion to enhance transformer capabilities. Depthwise separable convolutions combined with transformer insights enhance significant channels and features. The ACTB also includes a Context Fusion Block with a Redundant Information Processing Unit to filter redundant information and fuse local and global contexts. Extensive experiments on benchmark datasets with scale factors ×2, ×3, and ×4 show that CACN achieves state-of-the-art results with improvements in PSNR, SSIM, and LPIPS while maintaining lower computational overhead. These results demonstrate CACN’s ability to reconstruct high-resolution images efficiently.