一种高效的光场焦点堆栈的超分辨率网络OA

Light field(LF)cameras capture scenes from multiple perspectives,providing comprehensive image information.However,the resulting data often suffers from low image quality,and current research generally exhibits limitations in underutilizing multi-focus characteristics.The insufficient exploitation of deep feature capabilities directly constrains the visual enhancement effects of LF imaging.To address this issues,converting sub-aperture images(SAIs)is proposed into a focal stack(FS)using depth information and analyzing them.Leveraging the varying depth-of-field characteristics of FS,depth maps are generated through specifically designed filters to classify images as focused or defocused.A selective super-resolution(SR)network is then applied to enhance focused images,where a dimensionality stretching strategy is employed to integrate low-resolution FS with degradation maps corresponding to deterioration parameters as input to convolutional neural networks.Experiments were conducted on both public and self-built light field datasets.The results show that the proposed method has an average increase of 0.21%in the image information entropy for the foreground and 0.22%for the background compared to the LF-IINet method,and an average increase of 0.21%for the foreground and 0.28%for the background compared to the Distg-SSR method.In terms of the high-frequency energy of the image,the proposed method has an average increase of 10.78%for the foreground and 10.56%for the background compared to the LF-IINet method,and an average increase of 11.33%for the foreground and 11.02%for the background compared to the Distg-SSR method.Through depth-guided focused enhancement and defocused blurring strategies,this approach achieves a balance between detail enhancement and noise suppression.The framework provides an efficient and reliable solution for advancing LF imaging applications in computational photography.