基于零样本学习的核电站水下视频去噪方法OA

To address severe video degradation caused by high-energy particle radiation in real underwater environments of nuclear power plants,this study proposed a zero-shot underwater video denoising method that reduced the dependence on paired clean samples and task-specific training data,and improved the us-ability of visual information for underwater inspection.A pre-trained image restoration diffusion model was used as the backbone.Low-quality noisy videos were processed in batches,and key frames were selected as reference inputs.In the early denoising stage,hierarchical latent deformation and optical-flow alignment were used to impose global structural constraints and local motion constraints.In the middle and later stag-es,a hybrid flow-guided spatial-aware token merging strategy was introduced.Optical-flow matching was used in the downsampling stage to maintain motion consistency,spatially constrained similarity was used in the upsampling stage to preserve structural details,and a merging-ratio annealing mechanism was adopt-ed to dynamically balance temporal consistency and detail fidelity.On the collected underwater noisy video dataset from a nuclear power plant,the proposed method achieved MaxVQA,CLIPVQA,and Ewarp val-ues of 34.2,0.97,and 1.2,respectively,outperforming the compared video restoration and denoising methods.The proposed method enables effective video denoising in underwater nuclear power plant sce-narios where annotated training data and clean reference videos are difficult to obtain,and provides techni-cal support for underwater optical imaging enhancement,intelligent image analysis,and equipment defect detection.