办公室噪声语料数据集

Data Set Information: AIMS AND PURPOSES This corpus is intended to do cleaning (or binarization) and enhancement of noisy grayscale printed text images using supervised learning methods. To this end, noisy images and their corresponding cleaned or binarized ground truth are provided. Double resolution ground truth images are also provided in order to test superresolution methods. CORPUS DIRECTORIES STRUCTURE SimulatedNoisyOffice folder has been prepared for training, validation and test of supervised methods. RealNoisyOffice folder is provided for subjective evaluation. . |-- RealNoisyOffice | |-- real_noisy_images_grayscale | `-- real_noisy_images_grayscale_doubleresolution `-- SimulatedNoisyOffice |-- clean_images_binaryscale |-- clean_images_grayscale |-- clean_images_grayscale_doubleresolution `-- simulated_noisy_images_grayscale RealNoisyOffice - real_noisy_images_grayscale: 72 grayscale images of scanned 'noisy' images. - real_noisy_images_grayscale_doubleresolution: idem, double resolution. SimulatedNoisyOffice - simulated_noisy_images_grayscale: 72 grayscale images of scanned 'simulated noisy' images for training, validation and test. - clean_images_grayscale_doubleresolution: Grayscale ground truth of the images with double resolution. - clean_images_grayscale: Grayscale ground truth of the images with smoothing on the borders (normal resolution). - clean_images_binary: Binary ground truth of the images (normal resolution). DEscriptION Every file is a printed text image following the pattern FontABC_NoiseD_EE.png: A) Size of the font: footnote size (f), normal size (n) o large size (L). B) Font type: typewriter (t), sans serif (s) or roman (r). C) Yes/no emphasized font (e/m). D) Type of noise: folded sheets (Noise f), wrinkled sheets (Noise w), coffee stains (Noise c), and footprints (Noise p). E) Data set partition: training (TR), validation (VA), test (TE), real (RE). For each type of font, one type of Noise: 17 files * 4 types of noise = 72 images. OTHER INFORMATION 200 ppi => normal resolution 400 ppi => double resolution Attribute Information: The format of each file is the following: Grayscale PNG files ([Web link]). The ground truth is also provided as grayscale PNG files, and for the binary version the values are saturated to 0 and 255. Relevant Papers: J.?L.?Adelantado-Torres,?J.?Pastor-Pellicer,?and?M.?J.?Castro-Bleda.?Una?aplicación?móvil?para?la?captura?y mejora?de?imágenes?de?textos,?in:?V?Jornadas?TIMM?(Tratamiento?de?la?Información?Multilingüe?y Multimodal),?Red?temática?TIMM?(Tratamiento?de?Información?Multilingüe?y?Multimodal),?Sevilla,?2014.? M.?J.?Castro-Bleda,?S.?Espa?a-Boquera?and?F.?Zamora-Martinez.?Encyclopedia?of?Artificial?Intelligence, chapter?Behaviour-based?Clustering?of?Neural?Networks,?pages?144-151,?Information?Science?Reference, 2009.? F.?Zamora-Martinez,?S.?Espa?a-Boquera?and?M.?J.?Castro-Bleda.?Behaviour-based?Clustering?of?Neural Networks?applied?to?document?Enhancement,?in:?Computational?and?Ambient?Intelligence,?pages?144-151, Springer,?2007. Citation Request: Please refer to the Machine Learning Repository's citation policy [Web link]. For the database: F. Zamora-Martinez, S. Espa?a-Boquera and M. J. Castro-Bleda, Behaviour-based Clustering of Neural Networks applied to document Enhancement, in: Computational and Ambient Intelligence, pages 144-151, Springer, 2007. M.J. Castro-Bleda (1), S. Espa?a-Boquera (1), J. Pastor-Pellicer (1), F. Zamora-Martinez (2) mcastro '@' dsic.upv.es, sespana '@' dsic.upv.es, jpastor '@' dsic.upv.es, francisco.zamora '@' uch.ceu.es (1) Departamento de Sistemas Informáticos? y Computación, Universitat Politècnica? de València, Valencia, Spain (2) Departamento de Ciencias Físicas, Matemáticas y de la Computación, Universidad CEU Cardenal Herrera, Alfara del Patriarca, València, Spain

数据集信息：目标与用途
本语料库旨在通过监督学习方法，对带有噪声的灰度印刷文本图像进行清理（或二值化）与增强。为此，本数据集提供了带噪声图像及其对应的清理后/二值化基准真值（ground truth）图像；同时还提供了双分辨率基准真值图像，用于测试超分辨率（superresolution）方法。

语料库目录结构
`SimulatedNoisyOffice` 文件夹用于监督学习方法的训练、验证与测试；`RealNoisyOffice` 文件夹用于主观评估。
.
|-- RealNoisyOffice
| |-- real_noisy_images_grayscale
| `-- real_noisy_images_grayscale_doubleresolution
`-- SimulatedNoisyOffice
|-- clean_images_binary
|-- clean_images_grayscale
|-- clean_images_grayscale_doubleresolution
`-- simulated_noisy_images_grayscale

### RealNoisyOffice
- real_noisy_images_grayscale：72张扫描得到的带噪声灰度图像。
- real_noisy_images_grayscale_doubleresolution：与上述图像一致，但分辨率为双份。

### SimulatedNoisyOffice
- simulated_noisy_images_grayscale：72张用于训练、验证与测试的扫描得到的模拟带噪声灰度图像。
- clean_images_grayscale_doubleresolution：双分辨率灰度基准真值图像。
- clean_images_grayscale：常规分辨率的灰度基准真值图像，其边界经过平滑处理。
- clean_images_binary：常规分辨率的二值化基准真值图像。

文件命名规则
所有文件均为遵循`FontABC_NoiseD_EE.png`格式的印刷文本图像，各字段含义如下：
A) 字体大小：脚注尺寸（f）、常规尺寸（n）或大尺寸（L）；
B) 字体类型：打字机字体（t）、无衬线字体（s）或衬线罗马字体（r）；
C) 字体是否强调：是（e）/否（m）；
D) 噪声类型：纸张折叠褶皱（Noise f）、纸张起皱（Noise w）、咖啡污渍（Noise c）以及脚印污渍（Noise p）；
E) 数据集分区：训练集（TR）、验证集（VA）、测试集（TE）以及真实噪声集（RE）。

每种字体配置对应4种噪声类型，共计17种字体配置×4种噪声类型=72张图像。

其他信息
- 200 ppi（像素每英寸，pixels per inch）为常规分辨率；
- 400 ppi为双分辨率。

属性信息
每个文件的格式为灰度PNG（Portable Network Graphics）图像。基准真值图像同样以灰度PNG格式提供，其中二值化版本的像素值被饱和至0和255。

相关论文
1. J. L. Adelantado-Torres、J. Pastor-Pellicer 与 M. J. Castro-Bleda. 一款用于文本图像采集与增强的移动应用[C]//第V届多语言与多模态信息处理研讨会（TIMM）论文集，多语言与多模态信息处理主题网络，塞维利亚，2014。
2. M. J. Castro-Bleda、S. España-Boquera 与 F. Zamora-Martinez. 基于行为的神经网络聚类[M]//《人工智能百科全书》，第144-151页，信息科学参考出版社，2009。
3. F. Zamora-Martinez、S. España-Boquera 与 M. J. Castro-Bleda. 基于行为的神经网络聚类在文档增强中的应用[C]//计算与环境智能，第144-151页，Springer出版社，2007。

引用要求
请遵循机器学习仓库的引用规范[Web link]。针对本数据集的引用格式如下：
F. Zamora-Martinez、S. España-Boquera 与 M. J. Castro-Bleda. 基于行为的神经网络聚类在文档增强中的应用[C]//计算与环境智能，第144-151页，Springer出版社，2007。

作者及联系方式
M.J. Castro-Bleda (1)、S. España-Boquera (1)、J. Pastor-Pellicer (1)、F. Zamora-Martinez (2)
邮箱：mcastro '@' dsic.upv.es、sespana '@' dsic.upv.es、jpastor '@' dsic.upv.es、francisco.zamora '@' uch.ceu.es
(1) 西班牙巴伦西亚理工大学信息系统与计算系，巴伦西亚
(2) 西班牙CEU卡德纳尔·埃雷拉大学物理、数学与计算机科学系，阿尔法拉德尔帕特里亚尔，巴伦西亚