Enhancing angular resolution using layers obtained from light field superpixel segmentation

Synthesizing light field views based on superpixels in epipolar plane image domain

“…For PSNR and SSIM of each LF, the metrics are averaged over the synthesized views. We observe that our proposed approach achieves higher PNSR and SSIM values compared to that in [4]. However, total time consumption was larger than [4], which is mainly due to the total number of superpixels exploited in our framework being larger than that in [4], thereby letting more superpixels to be processed during the view-synthesis process.…”

“…To render a lower-resolution LF, we made a sparse set of 5 × 5 views by choosing the views with odd indices from the input LFs. We compared our method against our previous work [4]. In the case of [4], LF images are segmented with various superpixel size parameter ranging from 20~45.…”

“…We compared our method against our previous work [4]. In the case of [4], LF images are segmented with various superpixel size parameter ranging from 20~45. When comparing with our method, we chose the result of view-synthesis with superpixel size being equal to 45 since it was reported in [4] to have the best result both in terms of reconstruction quality and the execution time.…”

“…In the case of [4], LF images are segmented with various superpixel size parameter ranging from 20~45. When comparing with our method, we chose the result of view-synthesis with superpixel size being equal to 45 since it was reported in [4] to have the best result both in terms of reconstruction quality and the execution time. In our method, we did not vary the size of superpixel like [4] since the process in [2] to segment EPI which consists of line detection and line pairing gives only single possible segmentation result.…”

“…Yim and et al [4] proposed an LF view synthesis scheme which first segments the LF into LF superpixels by grouping similar pixels within and across all views using framework in [8], and subsequently shifts the superpixels by the amount of the disparity to reconstruct the targeting view. However, the method left lots of artifacts and produced noisy images.…”