1. What have the authors contributed in "Edge-based blur kernel estimation using patch priors" ?
In this paper the authors introduce a new patch-based strategy for kernel estimation in blind deconvolution.. Their approach estimates a “ trusted ” subset of x by imposing a patch prior specifically tailored towards modeling the appearance of image edge and corner primitives.. To choose proper patch priors the authors examine both statistical priors learned from a natural image dataset and a simple patch prior from synthetic structures.. The authors show that their patch prior prefers sharp image content to blurry ones.. A comprehensive evaluation shows that their approach achieves state-of-theart results for uniformly blurred images.
read more
2. What have the authors stated for future works in "Edge-based blur kernel estimation using patch priors" ?
As future work, the authors would like to extend their image formation model to handle more severe noise and other outliers to make it more robust on low quality inputs.
read more
3. How do the authors improve the performance of their image formation model?
As future work, the authors would like to extend their image formation model to handle more severe noise and other outliers to make it more robust on low quality inputs.
read more
4. What is the way to learn a patch prior?
The authors would like their patch prior to be sufficiently expressive, i.e., any edge patch P in natural images can be approximated by P = σZ+µ+ , where σ is the patch contrast, µ is the patch intensity, is a small error term.
read more
![Table 1. Quantitative comparison for each method: mean PSNR, mean SSIM, and geometric mean for error ratio, computed over the 32 test images from [13, 14].](/figures/table-1-quantitative-comparison-for-each-method-mean-psnr-2bysqwzg.webp)
![Figure 5. Comparison of results on one test image from [13]. Our kernels are less noisy and better resemble the ground truth (leftmost column). Sparse deconvolution with identical parameters are applied to all compared methods except Cho and Lee [4].](/figures/figure-5-comparison-of-results-on-one-test-image-from-13-our-d6teuxjn.webp)
![Figure 6. Performance comparison using the error ratio measure as in Levin et al. [13, 14]. The geometric mean of error ratios is shown in the legend for each algorithm. A ratio larger than 3 is deemed visually unacceptable, whereas a ratio less than 1 means the estimated kernel can do better than the ground truth kernel under the given sparse deconvolution method. It is worth mentioning that several estimated kernels from [4] appear better than the ground truth kernels. This could be due to the fact that Cho and Lee used a different deconvolution method to produce the final latent image x.](/figures/figure-6-performance-comparison-using-the-error-ratio-3hv9frlg.webp)
![Figure 7. Performance on our synthetic test set of 640 images. Top: comparison of success rate vs error ratio for all competiting methods. Our methods (thicker lines) significantly outperform all others on this test set. Bottom: a bar plot for the success rates at an error ratio of 3, which is deemed as the threshold for visually plausible deblurred results by Levin et al. [14].](/figures/figure-7-performance-on-our-synthetic-test-set-of-640-images-3rr2fcg4.webp)
![Figure 1. Algorithm pipeline. Our algorithm iterates between x-step and k-step with the help of a patch prior for edge refinement process. In particular, we coerce edges to become sharp and increase local contrast for edge patches. The blur kernel is then updated using the strong gradients from the restored latent image. After kernel estimation, the method of [20] is used for final non-blind deconvolution.](/figures/figure-1-algorithm-pipeline-our-algorithm-iterates-between-x-1d9yjscx.webp)
