Matrix Completion with Application to Missing Pixels Reconstruction
We received many electro-optical (EO) images from the Office of Naval Research (ONR). The COTR is Dr. Wen Masters. The missing pixels in R, G, and B bands are totally random. We developed novel matrix completion algorithms to deal with missing data in images. Fig. 1 shows reconstructed images of 1 file with 6 missing rates. The reconstruction performance is excellent even in high missing rates (99% and 99.9%). In terms of peak signal-to-noise ratio (PSNR), the scores are all above 25 dB. The visual appearance of reconstructed image in 99.9% missing data case is quite good, as we can still infer structural information out of the images.
 |
 |
 |
| 50% missing; PSNR: 45.84 dB | 75% missing; PSNR: 41.60 dB | 90% missing; PSNR: 37.10 dB |
 |
 |
 |
| 95% missing; PSNR: 33.83 dB | 99% missing; PSNR: 29.93 dB | 99.9% missing; PSNR: 26.32 dB |