Image region dividing apparatus
First Claim
1. An image region dividing apparatus comprising:
- image input means for inputting a mixed image in which character information and picture/graphics information are mixed as a digital image;
same-kind image region extraction means for;
selecting same kinds of the character information and the picture/graphics information in the input image by comparison,dividing the selected information into small region images of a size including the selected information,grouping the small region images into the same kinds of information,obtaining a position and size of each of the small region images, andspecifying a region in which it is discriminated whether the small region image is character information or picture/graphics information;
local feature pattern detection means for detecting local feature patterns formed by mutually adjacent pixels of the small region images, for each of the small region images grouped in the image;
means for calculating a frequency distribution of the local feature patterns;
correction means for;
obtaining a first classification vector based on each of differences obtained from horizontal and vertical pixel alignments of the small region images, andprojecting the first classification vector to the frequency distribution of the local feature patterns to correct the frequency distribution;
frequency distribution normalization means for normalizing the frequency distribution corrected by the correction means using a random number;
image kind identification means for;
receiving the normalized frequency distribution andidentifying whether the information included in the small regions are character information or picture/graphics information; and
image kind determination means for determining image kinds on the basis of the identification results from said image kind identification means; and
said image kind identification means including;
block image storage means for temporarily storing the small region images from said same-kind image region extraction means;
histogram calculation means for calculating a gradient vector direction histogram and a corrected luminance vector histogram of the small region images read out from said block image storage means;
supervising data storage means for pre-storing a plurality of classification vectors which are used as supervising data and predetermined in units of the kinds of images based on the character information and picture/graphics information;
inner product calculation means for performing inner product calculations of the gradient vector direction histogram and the corrected luminance vector histogram, and the first classification vector, in units of images of the small regions output from the histogram calculation means; and
buffer means for;
temporarily storing calculation results of said inner product calculation means, andoutputting the calculation results to feature discrimination means comprising a neural network for discriminating the kinds of an image;
FST means having a second classification vector to be used for classification between a handwritten character and a graphics image, andsaid image kind discrimination means calculates;
inner products of the input small region images and the second classification vector, andinner products of the gradient vector direction histogram and the corrected luminance vector histogram, and the first classification vector calculated by the histogram calculation means; and
said image kind discrimination means discriminates image kinds on the basis of results from the inner products.
0 Assignments
0 Petitions
Accused Products
Abstract
An image region dividing apparatus includes a same-kind image region extraction unit for dividing a digital image into blocks by extracting boundaries, from the background, of regions where same kinds of images are present, from the digital image; horizontal and vertical difference detectors for obtaining the difference values of the luminance levels of adjacent pixels in the horizontal and vertical directions from a discrimination target block; a feature pattern discrimination unit for performing recognition processing on the basis of a correlation between the shapes of a calculated corrected luminance level histogram ys and a calculated gradient vector direction histogram θr; and an image kind determination unit for determining image kinds. The apparatus performs image kind discrimination of each discrimination target block.
138 Citations
10 Claims
-
1. An image region dividing apparatus comprising:
-
image input means for inputting a mixed image in which character information and picture/graphics information are mixed as a digital image; same-kind image region extraction means for; selecting same kinds of the character information and the picture/graphics information in the input image by comparison, dividing the selected information into small region images of a size including the selected information, grouping the small region images into the same kinds of information, obtaining a position and size of each of the small region images, and specifying a region in which it is discriminated whether the small region image is character information or picture/graphics information; local feature pattern detection means for detecting local feature patterns formed by mutually adjacent pixels of the small region images, for each of the small region images grouped in the image; means for calculating a frequency distribution of the local feature patterns; correction means for; obtaining a first classification vector based on each of differences obtained from horizontal and vertical pixel alignments of the small region images, and projecting the first classification vector to the frequency distribution of the local feature patterns to correct the frequency distribution; frequency distribution normalization means for normalizing the frequency distribution corrected by the correction means using a random number; image kind identification means for; receiving the normalized frequency distribution and identifying whether the information included in the small regions are character information or picture/graphics information; and image kind determination means for determining image kinds on the basis of the identification results from said image kind identification means; and said image kind identification means including; block image storage means for temporarily storing the small region images from said same-kind image region extraction means; histogram calculation means for calculating a gradient vector direction histogram and a corrected luminance vector histogram of the small region images read out from said block image storage means; supervising data storage means for pre-storing a plurality of classification vectors which are used as supervising data and predetermined in units of the kinds of images based on the character information and picture/graphics information; inner product calculation means for performing inner product calculations of the gradient vector direction histogram and the corrected luminance vector histogram, and the first classification vector, in units of images of the small regions output from the histogram calculation means; and buffer means for; temporarily storing calculation results of said inner product calculation means, and outputting the calculation results to feature discrimination means comprising a neural network for discriminating the kinds of an image; FST means having a second classification vector to be used for classification between a handwritten character and a graphics image, and said image kind discrimination means calculates; inner products of the input small region images and the second classification vector, and inner products of the gradient vector direction histogram and the corrected luminance vector histogram, and the first classification vector calculated by the histogram calculation means; and said image kind discrimination means discriminates image kinds on the basis of results from the inner products.
-
-
2. An image region dividing apparatus comprising:
-
image input means for inputting a mixed image in which character information and picture/graphics information are mixed as a digital image; same-kind image region extraction means for; dividing the input digital image into blocks which can be differentiated from each other in terms of information, categorizing the blocks into groups by feature pattern, setting regions by fusing blocks together, and extracting a size of a region and an address thereof which indicates a position in the image; and image kind identification means for; obtaining a vector from a difference in pixel arrangement in an image of the region, identifying a feature pattern from said vector by a neural network learned in advance, and determining an image kind from the feature pattern. - View Dependent Claims (3, 4, 5, 6, 7, 8, 9)
-
-
10. An image region dividing apparatus comprising:
-
image input means for inputting a mixed image in which character information and picture/graphics information are mixed as a digital image; same-kind image region extraction means for; selecting same kinds of the character information and the picture/graphics information in the input image by comparison, dividing the selected information into small region images of a size including the selected information, grouping the small region images into the same kinds of information, obtaining a position and size of each of the small region images, and specifying a region in which it is discriminated whether the small region image is character information or picture/graphics information; local feature pattern detection means for detecting local feature patterns formed by image pixels of an image adjacent to the small region images, for each of the small region images grouped in the image; means for calculating a frequency distribution of the local feature patterns; correction means for; obtaining a first classification vector based on each of differences obtained from horizontal and vertical pixel alignments of the small region images, and projecting the first classification vector to the frequency distribution of the local feature patterns to correct the frequency distribution; frequency distribution normalization means for normalizing the frequency distribution corrected by the correction means using a random number; image kind identification means for; receiving the normalized frequency distribution, and identifying whether the information included in the small regions are character information or picture/graphics information; image kind determination means for determining image kinds on the basis of the identification results from said image kind identification means; image compression means for compressing the mixed image in which the input character information and picture/graphics information are mixed; wherein said apparatus outputs natural image, compressed graphics images or compressed character images mainly including compressed photographs, on the basis of the kind of the small region images determined by the image kind determination means; and said image compression means comprises; compression means for compressing the input image to a 1/4 reduced image; first layer data generation means for generating first layer data by further compressing the 1/4 reduced image compressed by said compression means to another 1/4 reduced image to produce a 1/16 reduced image; enlargement means for enlarging the first layer data (1/16 reduced image) two-fold; difference image generation means for calculating a difference image by subtracting the 1/4 reduced image from the image enlarged by said enlargement means; second layer data generating means for generating second layer data by compressing the difference image to a 1/4 reduced difference image; difference image decoding means for generating a ×
2 image by decoding the second layer data, and adding the ×
2 image to the image enlarged by said enlargement means to enlarge the ×
2 image two-fold; andthird layer data generation means for generating third layer data by adding the difference image obtained by said difference image decoding means to the input image; and said image kind determination means performs image kind discrimination in response to inputting the first, second, and third layer data to said image kind identification means.
-
Specification