Class-Adapted Blind Deblurring of Document Images

Ljubenovic, Marina; Zhuang, Lina; Figueiredo, Mário A. T.

doi:10.1109/icdar.2017.123

Cited by 11 publications

(11 citation statements)

References 28 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Namely, we introduce the patch classification step, yielding a simultaneous segmentation/deblurring method, which exploits the synergy between these two tasks: by identifying the most likely type of contents of each patch, the most adequate denoiser is used at that location, while a better deblurred image facilitates segmentation. Additionally, the experimental results reported in this paper considerably exceed those reported in [27] and [28].…”

Section: Related Work and Contributionscontrasting

confidence: 83%

“…That approach allows handling situations where the image being processed contains regions of different classes, as done by Teodoro et al for denoising and non-blind deblurring [44]. Additionally, we show that a similar framework with a dictionary-based prior can be used for BID when multiple classes are present in the same image [28]. Anwar et al exploit the potential of class-specific image priors for recovering spatial frequencies attenuated by the blurring process, by capturing the Fourier magnitude spectrum of an image class across all frequency bands [7] and that method achieves state-of-the-art results for images that belong to some specific classes (e.g., faces, animals, common objects), but there is no straightforward extension of the method for images that contain two or more classes.…”

Section: Related Work and Contributionsmentioning

confidence: 75%

“…That method is suitable for images that belong to a specific class (text, faces, or fingerprints) and a variety of blurring filters. Furthermore, in [28], we show how the previous method can be extended to tackle images that contain two classes and introduce different classspecific denoisers based on dictionaries trained from a clean image set. This paper combines and significantly extends the previous work reported in [27] and [28].…”

Section: Related Work and Contributionsmentioning

confidence: 99%

“…Furthermore, in [28], we show how the previous method can be extended to tackle images that contain two classes and introduce different classspecific denoisers based on dictionaries trained from a clean image set. This paper combines and significantly extends the previous work reported in [27] and [28]. Namely, we introduce the patch classification step, yielding a simultaneous segmentation/deblurring method, which exploits the synergy between these two tasks: by identifying the most likely type of contents of each patch, the most adequate denoiser is used at that location, while a better deblurred image facilitates segmentation.…”

Section: Related Work and Contributionsmentioning

confidence: 99%

“…In particular, we use denoisers based on Gaussian mixture models (GMM) or dictionaries, learned from patches of clean images from specific classes. These types of denoisers can be used for images that contain only one class, as presented in [27], or two or more classes [28].…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Plug-and-play approach to class-adapted blind image deblurring

Ljubenovic

Figueiredo

2019

IJDAR

View full text Add to dashboard Cite

Most of the existing single-image blind deblurring methods are tailored for natural images. However, in many important applications (e.g., document analysis, forensics), the image being recovered belongs to a specific class (e.g., text, faces, fingerprints) or contains two or more classes. To deal with these images, we propose a class-adapted blind deblurring framework, based on the plug-and-play scheme, which allows forward models of imaging systems to be combined with state-of-the-art denoisers. We consider three patch-based denoisers, two suitable for images that belong to a specific class and a general purpose one. Additionally, for images with two or more classes, we propose two approaches: a direct one, and one that uses a patch classification step before denoising. The proposed deblurring framework includes two priors on the blurring filter: a sparsity-inducing prior, suitable for motion blur and a weak prior, for a variety of filters. The results show the state-of-the-art performance of the proposed framework when applied to images that belong to a specific class (text, face, fingerprints), or contain two classes (text and face). For images with two classes, we show that the deblurring performance is improved by using the classification step. For these images, we choose to test one instance of the proposed framework suitable for text and faces, which is a natural test ground for the proposed framework. With the proper (dictionary and/or classifier) learning procedure, the framework can be adapted to other problems. For text images, we show that, in most cases, the proposed deblurring framework improves OCR accuracy. Keywords Blind deblurring • Class-specific image priors • Plug-and-play • ADMM • Patch-based image processing The research leading to these results has received funding

show abstract

Section: Related Work and Contributionscontrasting

confidence: 83%

Section: Related Work and Contributionsmentioning

confidence: 75%

Section: Related Work and Contributionsmentioning

confidence: 99%

Section: Related Work and Contributionsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Plug-and-play approach to class-adapted blind image deblurring

Ljubenovic

Figueiredo

2019

IJDAR

View full text Add to dashboard Cite

show abstract

Learning to Clean: A GAN Perspective

Sharma¹,

Verma²,

Vig³

2019

Lecture Notes in Computer Science

View full text Add to dashboard Cite

In the big data era, the impetus to digitize the vast reservoirs of data trapped in unstructured scanned documents such as invoices, bank documents, courier receipts and contracts has gained fresh momentum. The scanning process often results in the introduction of artifacts such as salt-and-pepper / background noise, blur due to camera motion or shake, watermarkings, coffee stains, wrinkles, or faded text. These artifacts pose many readability challenges to current text recognition algorithms and significantly degrade their performance. Existing learning based denoising techniques require a dataset comprising of noisy documents paired with cleaned versions of the same document. In such scenarios, a model can be trained to generate clean documents from noisy versions. However, very often in the real world such a paired dataset is not available, and all we have for training our denoising model are unpaired sets of noisy and clean images. This paper explores the use of Generative Adversarial Networks (GAN) to generate denoised versions of the noisy documents. In particular, where paired information is available, we formulate the problem as an image-to-image translation task i.e, translating a document from noisy domain ( i.e., background noise, blurred, faded, watermarked ) to a target clean document using Generative Adversarial Networks (GAN). However, in the absence of paired images for training, we employed CycleGAN which is known to learn a mapping between the distributions of the noisy images to the denoised images using unpaired data to achieve image-to-image translation for cleaning the noisy documents. We compare the performance of CycleGAN for document cleaning tasks using unpaired images with a Conditional GAN trained on paired data from the same dataset. Experiments were performed on a public document dataset on which different types of noise were artificially induced, results demonstrate that CycleGAN learns a more robust mapping from the space of noisy to clean documents.

show abstract