用于物料混合均匀性检测的高光谱图像散焦模糊去除OA

Detection of material mixing uniformity is critical for enabling online quality monitoring and pro-cess optimization.This study addresses the degradation of uniformity evaluation caused by defocus blur in hyperspectral imaging(HSI).A physics-constrained self-supervised learning framework for unpaired hy-perspectral image deblurring(PC-SSL-HSI)is proposed.A Uformer-based architecture incorporating the SimAM attention mechanism is employed as the deblurring network,while adversarial training is intro-duced to align deblurred outputs with clear images in the feature space.In addition,a blur kernel predic-tion module is designed based on a classical degradation model to construct pseudo-sample pairs,enabling self-supervised learning that guides the network to emphasize local detail restoration in hyperspectral imag-es.Experimental results demonstrate that the proposed method effectively enhances image detail,suppress-es artifacts,and improves the accuracy of material mixing uniformity evaluation.On a simulated dataset,the peak signal-to-noise ratio(PSNR)reaches 34.970 and the structural similarity index(SSIM)reaches 0.900,with concentration prediction errors ranging from 0.022 8 to 0.031 2.Furthermore,hyperspectral imaging experiments for material mixing uniformity indicate that the proposed method outperforms compar-ative approaches in metrics such as Kullback-Leibler divergence and coefficient of variation,highlighting its strong potential for engineering applications.