基于多尺度特征与扩散模型的高光谱图像重建OA

To improve the spatial resolution and quality of hyperspectral images,this paper proposes a super-resolution reconstruction model based on multi-scale feature extraction and a diffusion model.The model adopts an encoder-decoder architecture with a dual-stream residual encoder and a diffusion-enhanced network.The encoder uses vision transformer(ViT)and VMamba to establish the spectral global correlation model and spatial dynamic system,respectively.In the decoding stage,a deterministic ODE solving framework combined with score matching is adopted to accurately restore high-frequency de-tails.A high-frequency guided cross-attention mechanism is used to selectively enhance high-frequency features.Experimental results on the Indian Pines dataset show that compared with Deep HS,the pro-posed algorithm increases SRE by 5.12%,MPSNR by 4.18%,and MSSIM by 1.68%,while reducing SAM by 1.89%.On the Pavia University dataset,the corresponding improvements are 5.48%for SRE,3.01%for MPSNR,0.86%for MSSIM,and a reduction of 2.43%for SAM,which verifies the effective-ness of the proposed model.