融合小波残差与稀疏先验的无人机运动图像去模糊方法OA

UAV aerial target detection technology is an important support for smart grid operation and maintenance.However,due to lens jitter or high-speed camera movement,images captured by UAVs often suffer from motion blur.Current mainstream deep learning models face challenges in training because of the lack of datasets for UAV motion blur.To address this issue,a mapping model with blur synthesis characteristics is adopted to solve the problem of dataset scarcity.A wavelet residual block is introduced and embedded in the downsampling stage of the model's codec.Wavelet transform is used to process features from the spatial domain to the wavelet domain,avoiding frequency information loss due to downsampling.Additionally,a sparse prior model is incorporated to generate initial values for the DIP rather than using random initial values.This approach utilizes prior mechanism knowledge to guide the network out of local optima,enhancing the quality of clear images.Experimental results demonstrate that:(1)The generated blurred images are highly similar to real blurred images and can be directly used for model training;(2)Compared with current deep learning-based deblurring algorithms,the proposed algorithm exhibits superior performance and better deblurring effect;in comparison to the baseline model,the PSNR increases by 0.92 dB,while the SSIM improves by 0.024 2.This research not only provides an algorithm for the clarification of UAV aerial images,but also can be applied to similar image motion deblurring scenarios.