Sparse Document Image Coding for Restoration
read more
Citations
Resolution enhancement of textual images via multiple coupled dictionaries and adaptive sparse representation selection
Non-Local Sparse Image Inpainting for Document Bleed-Through Removal
A study about the reconstruction of remote, low resolution mobile captured text images for OCR
Handling noise in textual image resolution enhancement using online and offline learned dictionaries
Joint denoising and magnification of noisy Low-Resolution textual images
References
Compressed sensing
Matching pursuits with time-frequency dictionaries
$rm K$ -SVD: An Algorithm for Designing Overcomplete Dictionaries for Sparse Representation
Image Denoising Via Sparse and Redundant Representations Over Learned Dictionaries
Image Super-Resolution Via Sparse Representation
Related Papers (5)
A learning framework for degraded document image binarization using Markov Random Field
Frequently Asked Questions (19)
Q2. What contributions have the authors mentioned in the paper "Sparse document image coding for restoration" ?
In this paper, the authors explore the use of sparse representation based methods specifically to restore the degraded document images.
Q3. How do the authors avoid blocky artifacts in the reconstructed image?
In order to avoid blocky artifacts in the reconstructed image, the authors use overlapping patches for restoration and the final reconstructed image is obtained by performing averaging at the overlapped regions.
Q4. What techniques can be used to solve the above problem?
The techniques such as i) greedy methods (matching pursuit [12]) or ii) convex relaxation (l1-norm) can be used to solve the above problem.
Q5. How many times the size of the dictionary is fixed?
In order to maintain overcompleteness and recover sparse representation [5], size of dictionary is usually fixed to four times the size of the patch.
Q6. What is the way to represent a document image patch?
a document image patch (y) that the authors would like to represent using a dictionary D should be computed as:y = g(D,α), (3)where α is a set of parameters and g is a non-linear function that maps from the binary document dictionary elements to a valid binary document image or patch.
Q7. How long does it take to restore a word?
Their algorithm takes about 12 seconds to restore a document of size 157 × 663 on a 2GB RAM and Intel(R) Core(TM) i3−2120 system with 3.30 GHz processor with un-optimized implementation.
Q8. What are the main challenges in restoring document images?
(3) Noise in document images usually contain a mixture of degradations coming from independent processes such as erosion, cuts, bleeds, etc.
Q9. What is the fundamental assumption in a sparse coding technique?
One of the fundamental assumptions in such a representation is that the elements of the dictionary span the subspace of images of interest and that any linear combination of a sparse subset of dictionary elements is indeed a valid image.
Q10. What is the reason for the degraded page?
The recognition error on the degraded page was due to erosion and low printing quality, which might possibly confuse the OCR when the noise fills up the gap between two characters in a word.
Q11. What is the effect of texture-blending on the document?
texture-blending simulates effects such as textured paper or stained paper, and was produced by linearly blending the document with a texture image for various degrees of blending.
Q12. How can one obtain clean documents with any font?
with the advent of internet, one can obtain clean documents with any font easily e.g simple search of ‘gothic text’ will result in lot of high quality documents which can be used to restore gothic texts.
Q13. What is the critical challenge in restoration of document images?
The authors used the sparse coding technique proposed in [3] treating the missing pixels (cuts) as infinite noise and restored the image after learning a dictionary using large number of clean text and natural image patches.
Q14. What is the effect of patch size on the restoration of degraded images?
If the patch size is large, the dictionary elements may overfit the training data, resulting in reduced flexibility of degraded images that can be restored.
Q15. What is the threshold parameter for the restoration of a dictionary?
It is observed in [1], [13], [3] that very large dictionary leads to overfitting i.e, learnt atoms may correspond to individual patches instead of generalizing for large number of patches and very small dictionary leads to underfitting.
Q16. What are the basic elements that constitute the documents?
The fundamental elements that constitute the documents are strokes, curves, glyffs, etc. and their method automatically learns these elements.
Q17. What kind of degradation is seen in word sanguinary?
Another kind of degradation is fading resulting in near cuts as seen in character a in word sanguinary, which is restored with high resolution.
Q18. What is the sparse representation of the image patch?
i.e,x ≈ Dα s.t. ||α||0 L, (1)where α is the sparse representation of the image patch and ||.||0 is l0 pseudo-norm, which gives a measure of number of non-zero entries in a vector, and the constant L defines the required sparsity level.
Q19. What is the sparse representation of x?
sparse representation of x is recovered from y asα̂ = min α ||α||0 s.t ||y −Dα||2 ≤ , (2)where is constant and can be tuned according to the application at hand.