Single Image Reflection Removal Through Cascaded Refinement

Li, Chao; Yang, Yixiao; He, Kun; Lin, Stephen; Hopcroft, John E.

doi:10.1109/cvpr42600.2020.00362

Cited by 143 publications

(178 citation statements)

References 24 publications

Supporting

Mentioning

177

Contrasting

Order By: Relevance

“…Whereas here R and B both are varying across created image set. Comparison with recent methods ( [21,31] and [32]) on synthetic as well as real datasets validates the efficacy of the proposed method.…”

Section: Resultssupporting

confidence: 55%

“…Zhang et al [21] produce a dull effect in images in real as well as synthetic set. Also reflection is more dominant in [21,31] and [32], for instance in first test image, reflected human face is clearly visible in all the other methods except ours. Second image has reflection of curtain folds, which is quiet prevalent in others compared to ours.…”

Section: Resultsmentioning

confidence: 67%

“…There is a ghosting cue like effect in some of our images due to varying background, which can be improved by experimenting with CNN hyper parameters. [21] ERRNet [31] IBCLN [32] ours Fig. 4.…”

Section: Resultsmentioning

confidence: 99%

“…Resent method Zhang [21], ERRNet [31], and IBCLN [32] are chosen for comparison with the state of the art. Since all the existing reflection test sets accommodate only single image methods, we had to evaluate our model on our own test set.…”

Section: Resultsmentioning

confidence: 99%

“…4 (along with their SSIM and PSNR scores), Where reduction in pixel's intensity can be seen in the output images of [21] and ERRNet [31]. IBCLN [32] does not reduce the intensity of whole image but it fails to reduce the intensity of reflection. Problem arises with the real test set, where no aligned background is available.…”

Section: Resultsmentioning

confidence: 99%

See 4 more Smart Citations

Two-Image Approach to Reflection Removal with Deep Learning

Chaurasiya

Ganotra

2021

Communications in Computer and Information Science

View full text Add to dashboard Cite

Reflection in an image is not always desirable due to the loss of information. Conventional methods to remove reflection are based on priors that require certain conditions to be fulfilled. Recent advancements of deep learning in many fields have revolutionized these traditional approaches. Using input images more than one reduces the ill-posedness of the problem statement.Standard assumption in numerous methods assumes background is stationary and only reflection layer is varying. However, images at different angles have slightly different backgrounds. Considering this a new dataset is created where both reflection and background layer is varying. In this paper, a two-image based method with an end to end mapping between the observed images and background is presented. The key feature is the practicability of the method, wherein a sequence of images at slightly different angles can easily be captured using modern dual camera mobile devices. A combination of feature loss with MSE maintains the content and quality of the resultant image.

show abstract

Section: Resultssupporting

confidence: 55%