2018/01/12 by Reuben A. Farrugia, Farrugia, Reuben A., Christine Guillemot +1
Computer Science · Environmental Science · #Advanced Vision and Imaging #Advanced Image Processing Techniques #Remote Sensing in Agriculture
paper · pdf · doi:10.48550/arxiv.1801.04314
Light field imaging has recently known a regain of interest due to the\navailability of practical light field capturing systems that offer a wide range\nof applications in the field of computer vision. However, capturing\nhigh-resolution light fields remains technologically challenging since the\nincrease in angular resolution is often accompanied by a significant reduction\nin spatial resolution. This paper describes a learning-based spatial light\nfield super-resolution method that allows the restoration of the entire light\nfield with consistency across all sub-aperture images. The algorithm first uses\noptical flow to align the light field and then reduces its angular dimension\nusing low-rank approximation. We then consider the linearly independent columns\nof the resulting low-rank model as an embedding, which is restored using a deep\nconvolutional neural network (DCNN). The super-resolved embedding is then used\nto reconstruct the remaining sub-aperture images. The original disparities are\nrestored using inverse warping where missing pixels are approximated using a\nnovel light field inpainting algorithm. Experimental results show that the\nproposed method outperforms existing light field super-resolution algorithms,\nachieving PSNR gains of 0.23 dB over the second best performing method. This\nperformance can be further improved using iterative back-projection as a\npost-processing step.\n