基于多维注意力网络的图像超分辨率重建

打开文本图片集

关键词：超分辨率；扩散模型；多维注意力网络；部分卷积 中图分类号：TP391.4 文献标识码：A doi：10.37188/CJLCD.2025-0058 CSTR：32172.14.CJLCD.2025-0058

Abstract：Existing single-image super-resolution reconstruction methods based on difusion probabilistic model are deficient in spatial feature information extraction，failing to fully mine the relevant information， as well as redundancy in the computational process. In this paper，a single-image super-resolution reconstruction method incorporating multidimensional atention network is designed.First，multidimensional attention is proposed on the basis of the SRDiffdifusion model，which combines channel attention，selfattention and spatial atention to enhance the model's ability to capture features at different scales，so that moredetails and better global consistency can be retained at the same time when recovering high-resolution images. Second，PConv partial convolution is introduced to accurately extract the spatial features of the image，improve the quality of the super-resolution results，and significantly reduce the amount of computation，thus improving the operational eficiency of the model. Under the condition of magnification factor of 4，this paper's method is compared with other methods on five test sets，and the results show that the peak signal-to-noise ratio of this paper's method is improved by O.762 dB compared with the average of the other compared methods，and the structural similarity is improved by O.O82 compared with the average of the other compared methods.The proposed method in this paper possesses subjectively more delicate details and more excellent visual effects，objectively has higher peak SNR values and structural similarity values.

Key words：super-resolution；diffusion model；multidimensional attention network；partial convolution

1引言

图像超分辨率作为计算机视觉与图像处理领域的经典任务，其核心在于以低分辨率图像为模板，通过特定技术手段，重建出饱含丰富细节与清晰纹理的高分辨率图像。（剩余15905字）