多维度聚合Transformer的图像超分辨率重建

打开文本图片集

MDAT ：Multi-dimensional aggregation transformer for image super-resolution reconstruction

CHEN Qingjiang，CHEN Pengmin

（School of Science， Xian University of Architecture and Technology，Xi'an 71OO55，China） * Corresponding author，E-mail： 409316492@qq. com

Abstract：To address the limitations of restricted receptive-field scales and insufcient exploration of additional dimensional information in existing Transformer-based image super-resolution networks，this paper proposed a multi-dimensional aggregation transformer network. First，a multi-scale interaction modulation module was designed to extract multi-scale features from low-resolution images，enhancing the diversity of information flow.Second，a spatial-channel interaction module was integrated into transformer layers， employing four types of attntion mechanisms to fully extract key features and achieve efective feature fusion，thereby improving model performance. Third，a feature-reuse transformer module was proposed to explicitly model inter-layer feature relationships，enabling precise extraction and eficient reuse of impor tant features.Experimental results demonstrate that the proposed method outperforms existing state-ofthe-art algorithms on five benchmark datasets. Specifically，in super-resolution tasks with various magnification factors，it achieves an average improvement of O.26 dB in peak signal-to-noise ratio and O.002 4 in structural similarity index measure compared to Swin Transformer-based methods，producing clearer reconstruction results. These findings validate the efectiveness of the proposed approach and its strong potential for practical applications in image super-resolution tasks.

Key words： image super-resolution； transformer； attention mechanism； feature interaction； feature re-use；multi-scale

1引言

单幅图像超分辨率（Single ImageSuper-Resolution，SISR）是一项经典的低级视觉任务，在医学成像、数字摄影和降低图片传输成本等多个领域应用广泛，其目标是将低分辨率（Low-Resolution，LR）图像重建为细节更加丰富的高分辨率（High-Resolution，HR）图像。（剩余19137字）