Identifying picture areas based on gradient image analysis
View Patent ↗In one embodiment, a method for identifying areas in a document image is provided. The method comprises generating binarized and gradient images based on the document image; and performing a classification operation to classify areas in the document image into one of a noise area and a picture area based on attributes computed on the binarized and gradient images.
1. A method for identifying areas in a document image, the method comprising:
generating a binarized document image based on the document image;
generating gradient images based on the document image, the gradient images comprising:
a first gradient image based on a horizontal component of a gradient of the document image,
a second gradient image based on a vertical component of the gradient of the document image, and
a third gradient image being a combination of said first gradient image and said second gradient image,
preliminary segmenting the binarized document image into areas; and
classifying, prior to a final segmentation of the document image, the preliminary segmented areas in the document image into classes based on attributes computed from the binarized document image, and the first, second and third gradient images, wherein each of the document classes is associated with a picture area, a text area, or a noise area, wherein the attributes comprise an average number of white holes in closed black contours on the third gradient image, wherein the white holes are white gaps with more than a threshold number of pixels that represent low contrast in brightness on the document image, and wherein the closed black contours represent high contrast in brightness on the document image.
2. The method of claim 1 , wherein the classifying is performed through the use of a plurality of classifiers.
3. The method of claim 1 , wherein the attributes comprise at least part of:
an average length of black vertical run-length encoding (RLE) strokes computed from the binarized image;
an average length of white vertical RLE strokes computed from the binarized image;
an average length of black horizontal RLE strokes computed from the binarized image; and
an average length of white horizontal RLE strokes computed from the binarized image.
4. The method of claim 1 , wherein the attributes comprise an average value, a value related to dispersion, a value related to a asymmetry coefficient, and a value related to a coefficient of excess computed on a horizontal gradient image and a vertical gradient image.
5. The method of claim 1 , wherein the classifying comprises combining the attributes and detecting the class based on the combined attributes.
6. The method of claim 5 , wherein detecting the class comprises:
computing a probability density for each possible class,
comparing the values of the probability densities, and
attributing the area to the class associated with the greatest probability density.
7. The method of claim 5 , wherein detecting the class for each area comprises:
computing a value associated with a probability density for each possible class for the respective area,
comparing the value associated with the probability density with an empirically determined constant, and
attributing the respective area to a class when the value associated with the probability density for the particular class is greater than the empirically determined constant.
8. The method of claim 1 , wherein the third gradient image is calculated for at least a separate area of the document image prior to the final segmentation.
9. The method of claim 1 , wherein the first, second, and third gradient images each comprise a grayscale image in which a brightness of each pixel corresponds to an absolute value of the respective or both horizontal or vertical component of the gradients of an initial grayscale image derived from the document image.
10. A document analysis system comprising:
a processor; and
a memory coupled to the processor, the memory storing instructions to perform a document analysis method for identifying areas in a document image, the method comprising:
generating a binarized document image based on the document image;
generating gradient images based on the document image, wherein the gradient images comprise:
a first gradient image based on a horizontal component of a gradient of the document image,
a second gradient image based on a vertical component of the gradient of the document image, and
a third gradient image being a combination of said first gradient image and said second gradient image;
preliminary segmenting the binarized document image into areas; and
classifying, prior to a final segmentation of the document image, the preliminary segmented areas in the document image into classes based on attributes computed from the binarized document image, and the first, second, and third gradient images, wherein each of the document class is associated with a picture area, a text area or a noise area, wherein the attributes comprise an average number of white holes in closed black contours on the third gradient image, and wherein the white holes are white gaps with more than a threshold number of pixels that represent low contrast in brightness on the document image, and the closed black contours represent high contrast in brightness on the document image.
11. The document analysis system of claim 10 , wherein the classifying includes the use of a plurality of classifiers.
12. The document analysis system of claim 10 , wherein the attributes comprise:
an average length of black vertical run-length encoding (RLE) strokes computed from the binarized image;
an average length of white vertical RLE strokes computed from the binarized image; an average length of black horizontal RLE strokes computed from the binarized image; and
an average length of white horizontal RLE strokes computed from the binarized image.
13. The document analysis system of claim 10 , wherein the attributes comprising an average value, a value related to dispersion, a value related to a asymmetry coefficient, and a value related to a coefficient of excess computed on the horizontal gradient image and vertical gradient image.
14. The document analysis system of claim 10 , wherein the third gradient image is calculated for at least a separate area of the document image prior to the final segmentation.
15. The document analysis system of claim 10 , wherein the first, second, and third gradient images each comprise a grayscale image in which a brightness of each pixel corresponds to an absolute value of the respective or both horizontal or vertical component of the gradients of an initial grayscale image derived from the document image.
16. A computer-readable non-transitory medium having stored thereon a sequence of instructions which when executed by a processing system cause the system to perform a method for identifying areas in a document image, the method comprising:
generating a binarized document image based on the document image;
generating gradient images based on the document image, wherein the gradient images comprise:
a first gradient image based on a horizontal component of a gradient of the document image,
a second gradient image based on a vertical component of the gradient of the document image, and
a third gradient image being a combination of said first gradient image and said second gradient image;
preliminary segmenting the binarized document image into areas, and
classifying, prior to a final segmentation of the document image, the preliminary segmented areas in the document image into classes based on attributes computed from the binarized document image, and the first, second and third gradient images, wherein each of the document class is associated with a picture area, a text area or a noise area, wherein the attributes comprise an average number of white holes in closed black contours on the third gradient image, wherein the white holes are white gaps with more than a threshold number of pixels that represent low contrast in brightness on the document image, and wherein the closed black contours represent high contrast in brightness on the document image.
17. The computer-readable non-transitory medium of claim 16 , wherein the classification operation uses a plurality of classifiers.
18. The computer-readable non-transitory medium of claim 16 , wherein the attributes comprise an average length of black and white vertical and horizontal Run-Length Encoding (RLE) strokes computed on the binarized image.
19. The computer-readable non-transitory medium of claim 16 , wherein the attributes are characterized in terms of average value, dispersion, asymmetry coefficient, and a coefficient of excess computed on the horizontal and vertical gradient images.
20. The method of claim 1 , wherein a gradient image further comprises an image generated such that a pixel of said image is a function of a brightness of a neighboring pixel.
21. The computer-readable non-transitory medium of claim 16 , wherein the third gradient image is calculated for at least a separate area of the document image prior to the final segmentation.
22. The computer-readable non-transitory medium of claim 16 , wherein the first, second, and third gradient images each comprises a grayscale image in which a brightness of each pixel corresponds to an absolute value of the respective or both horizontal or vertical component of the gradients of an initial grayscale image derived from the document image.