Light Field Image Angular Super-Resolution Using Edge Features

The existing Light Field (LF) angular super-resolution methods reconstruct angular high-resolution LF images with blurred local edge details and artifacts, due to the insufficient utilization of edge-features closely related to local edge-details. To this end, we improve and implement the LF image angular super-resolution method in terms of edge-feature extraction, utilization, and accuracy assurance. First, we use the Sobel operator to extract the spatial edge-magnitude image and angular edge-magnitude image of the LF image, and combine the neural network to extract complete edge-features from them. Secondly, we introduce a multi-level attention mechanism to perform weighted fusion of edge-features and spatial features and angular features containing global information, so as to achieve the integrity of global information while obtaining local edge details. Finally, we design a loss function based on edge-features which further supervises reconstructing local details and edges of the reconstructed image. Experimental results show that the PSNR values of the reconstructed LF images are improved by 0.25dB and 0.35dB respectively on real datasets and synthetic datasets compared with existing methods. The distortion of its local edge details is reduced, which makes up for the defect of image blur and becomes a factor in improving the PSNR value.