KR100412176B1 - Document segmentation compression, reconstruction system and method - Google Patents

Abstract

The present invention relates to a system and method for segmenting, compressing, and restoring a document including text and images. More particularly, the present invention relates to text and images in the case of a document composed of text and images. The present invention relates to a system and method for segmenting, compressing, and restoring a document by encoding and re-decoding each character by character and literally image by image.

According to the present invention, a document compression system includes a foreground image corresponding to a character and a background image corresponding to an image of a background and a picture, and a bitmap. A document splitting module for separating and generating an image; A JBIG encoder for generating a JBG file by receiving the bitmap image as an input and encoding the bitmap image; An FDWT unit which receives the foreground image and the background image as inputs and performs wavelet transform by a wavelet lifting method to output wavelet coefficients; And a bit plane encoder that receives the wavelet coefficient as an input and encodes the wavelet coefficient to generate a bitstream.

Description

Document segmentation compression, reconstruction system and method

The present invention relates to a system and method for segmenting, compressing, and restoring a document including text and images. More particularly, the present invention relates to text and images in the case of a document composed of text and images. The present invention relates to a segmentation, compression, and decompression system and method for document segmentation by encoding and re-decoding of characters into characters and images into images.

The most common compression method for still images is JPEG (Joint Photographic Experts Group) encoding.

Lossy compression methods, such as JPEG, increase the compression rate, resulting in higher loss of the original image.

In general, when the image is compressed using a lossy compression method such as JPEG, the reconstructed image can be easily seen by the human eye.

However, if a scanned image of a high resolution document such as a newspaper or a magazine, that is, a character is included and the image of a relatively high importance for the character is compressed by a lossy compression method such as JPEG, the image is not recognized. Hard.

This is because a character generally has an extremely small and fine area of an image compared to an image, so even a small loss of an image has a great effect on the character.

That is, in the past, text and images were compressed and restored together, and when compressing and restoring characters in the same way as compressing and restoring images, characters were often shuffled and blurred, making it difficult to read with eyes. .

In addition, conventionally, in performing an integer transform of the wavelet lifting method, 1) a character image including only a character, 2) a simple monochrome image, 3) a complex monochrome image, 4) a simple color image, 5) a natural image, Since only five cases were considered, the integer conversion could not be selected according to the type of image.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and before a document is compressed, a document and a document containing a character and an image are divided by encoding and decoding the characters and the images, respectively, and the images, respectively, as images. Can be compressed and reconstructed at the same time, and integer conversion can be selected to suit the input image, thereby reducing the loss. Therefore, the compression of documents containing characters and images that can maintain high compression and high quality of documents, It is an object of the present invention to provide a restoration system and method.

In order to achieve the above object, the present invention, a document for generating a document containing a character and an image, separated into a foreground image corresponding to the character, a background image corresponding to the image of the background and pictures, and a bitmap image A splitting module; A JBIG encoder for generating a JBG file by receiving the bitmap image as an input and encoding the bitmap image; An FDWT unit which receives the foreground image and the background image as inputs, performs color difference conversion and downsampling, and performs wavelet transform by a lifting method to output wavelet coefficients; Contains a character and an image including a bit-plane encoder that receives the wavelet coefficient as an input, performs graycode transformation and quantization, and generates a bitstream by performing encoding by a bitplane method. It is intended to provide a compression system for compiled documents.

In addition, the present invention includes a bit plane decoder for receiving a bit stream of the foreground image and the background image as an input to generate wavelet coefficients and perform inverse gray code conversion; A JBIG decoder which receives a bitstream of a JBG file as an input and generates a bitmap image; An IDWT unit which receives the wavelet coefficient as an input and performs inverse wavelet transform, upsampling, and color difference conversion to generate a foreground image corresponding to a character and a background image corresponding to the image; An object of the present invention is to provide a compression system for a document including text and images including a bitmap image, a foreground image, and a document reconstruction unit for restoring a background image to an original document.

In addition, the present invention, the first step of separating the entire image of the document containing the character and the image into a bitmap image, the foreground image of the character, and the background image of the image; Generating a JBG file by encoding the bitmap image; Performing a wavelet transform on the foreground image by a lifting method to output wavelet coefficients, and encoding the wavelet coefficients to generate a bitstream; Performing a wavelet transform on the background image by a lifting method to output wavelet coefficients, and encoding the wavelet coefficients to generate a bitstream; It is intended to provide a method of compressing a document that includes text and images to include.

The present invention also provides a first process for decoding a JBG file to generate a bitmap image; Decoding a bitstream of the foreground image to generate wavelet coefficients, and performing inverse wavelet transform on the wavelet coefficients to generate a foreground image; Generating a wavelet coefficient by decoding the bitstream of the background image and generating a background image by performing inverse wavelet transform on the wavelet coefficient; A fourth step of reconstructing and reconstructing the bitmap image, the foreground image, and the background image into an original document; An object of the present invention is to provide a method of restoring a document including text and images.

1 is a block diagram of a system for compressing a document according to the present invention.

2 is a block diagram of a system for restoring a document according to the present invention.

3 is a structure of a TSFB for performing FDWT upon compression of a document according to the present invention.

4 is a diagram illustrating an analysis process of TSFB according to the present invention.

5 is a diagram illustrating a one-dimensional reversible lifting transformation process for transformations according to the present invention.

FIG. 6 is a pseudo code part for performing lifting for each pixel in one dimension according to the present invention.

7 is a diagram illustrating a color difference signal downsampling process according to the present invention.

8 is a diagram illustrating a quantization process of wavelet coefficients according to the present invention.

9A and 9B illustrate a process of processing a bit plane and a subblock of a wavelet coefficient value according to the present invention.

10 is a diagram illustrating a synthesis process of TSFB according to the present invention.

11 is a diagram illustrating an image restoration process according to the present invention.

12A to 12C are original experimental images.

13A and 13B are compression ratios of a background image and a foreground image according to the present invention.

14A and 14B are PSNR values of a background image and a foreground image according to the present invention.

15A to 15C are reconstructed images after compressing and reconstructing the original experimental image.

<Description of Symbols for Main Parts of Drawings>

100: document splitting module 110: JBIG encoder

120,132: FDWT section 122,134: bit plane encoder

124,135: repeat length encoder 126,136: arithmetic encoder

210: JBIG decoder 220,230: Arithmetic decoder

222,232: Repeat length decoder 224,234: Bit plane decoder

226,236: IDWT unit 240: document reconstruction unit

Hereinafter, the present invention will be described with reference to the accompanying drawings.

1 is a block diagram of a system for compressing a document according to the present invention.

First, referring to the basic configuration of the sign system according to the present invention, a document containing a character and an image is generated by separating the foreground image corresponding to the character, the background image corresponding to the background and the image of the image, and the bitmap image. A document splitting module 100;

A JBIG encoder 110 which receives the bitmap image as an input and encodes the bitmap image to generate a JBG file;

An FDWT unit (120, 132) for receiving the foreground image and the background image as inputs and performing wavelet transformation by color difference conversion, downsampling, and lifting method to output wavelet coefficients;

Bit-plane encoders 122 and 134 which receive the wavelet coefficients as inputs, perform graycode transformation and quantization, and generate a bitstream by performing encoding by a bit-plane method.

In the compression of a document containing text and images, the entire image of the document is divided into a foreground area corresponding to a character and a background area corresponding to an image of a background and a picture, respectively. It is necessary to use a compression method, and the document splitting module 100 performs the task of separating the document into the foreground area, the background area, and the bitmap area.

The following is a compression process of the system according to the present invention.

(1) The input image is divided into a foreground area, a background area, and a bitmap.

(2) Color difference conversion is performed on the divided foreground area and the background area.

(3) Downsample the U and V values corresponding to the chrominance values in the chrominance-converted Y, U and V signals.

At this time, the downsampling ratio of the luminance value Y and the color difference values U and V is 4: 1: 1.

(4) Wavelet transform is performed on each of the Y, U, and V components.

(5) Gray code conversion is performed on the generated wavelet coefficients.

(6) Quantization is performed for each subband with respect to the wavelet coefficients converted into gray codes.

In this case, different quantizations are applied to the LL, LH, HL, and HH subbands.

(7) Next, a bit plane is formed for the quantized wavelet coefficients, and a subblock is formed in each bit plane.

(8) Subblock encoding is performed on each subblock in the bit plane.

(9) Iterative length coding and adaptive arithmetic coding are performed on the generated bitstream.

The entire image of the document including the text and the image is input to the document splitting module 100 as a .bmp file, a .jpg file, or a .tiff file.

12A, 12B, and 12C are images of the original document.

The document splitting module 100 performs color clustering based on the K-means algorithm.

Color clustering using the K-average algorithm is an algorithm for clustering by comparing the distance between all pixel values of an image and an average value of pixels belonging to each cluster when the color distribution of an image is placed on RGB coordinates. Initialize each cluster by the number of clusters to divide first.

In the present invention, the entire image is clustered in the foreground color and the background color, and color clustering using the K-average algorithm operates in the following order.

1. Reset the foreground color to black and the background color to white.

2. Compare all pixel values in the image with the distance between the foreground and background colors, and include them on the closer side.

3. The average of each pixel included in the foreground and the pixel included in the background is calculated and updated to the foreground color and the background color.

4. Repeat steps 2 and 3 until all colors are collected.

However, the image of a general document is rarely limited to two colors, and the design and contrast of the document cause a change in the background color and the foreground color throughout the image.

This algorithm is a block bicolor clustering algorithm that extends the K-means color clustering algorithm.

The block bicolor clustering algorithm applies a K-average color clustering algorithm by dividing an entire image into a plurality of square blocks, and can cope with a color change problem of the entire image.

The size of the square block should be small enough to detect the change in color of a character in a sentence.

In this case, the entire block of the small block may be located in the foreground area or the background area. In this case, the color value of the brightest pixel is set as the background color.

This method selects pixels in the foreground that are not in the foreground because of the nature of the continuous pixel values in blocks that do not contain either characters or images and that are very continuous. In order to cope with this, the algorithm that extends the block bicolor clustering algorithm is a multiscale bicolor clustering algorithm.

The multiscale bicolor clustering algorithm is performed by considering a continuous grid in which the block size becomes smaller instead of one block size.

1. The block bicolor clustering algorithm is applied by dividing the whole image into a grid of a certain block size.

2. Divide the entire image into a grid of block size smaller than the block size of the previous grid.

3. Initialize the foreground and background colors of each block to the foreground and background colors of the previous grid block to which the current blocks belong.

4. For each block, compare all pixel values in the block and the distances between the foreground and background colors, respectively, and include them on the closer side.

5. The foreground and background color values of the previous grid block are added to each of the pixels included in the foreground and the pixels included in the background, and the average value is updated to the foreground color and the background color.

6. Repeat 4 and 5 until each color is collected and repeat from 2 until it is the size of a block.

The multiscale bicolor clustering algorithm uses the foreground color and the background color of the previous block when initializing the block.

This color initialization process creates a tendency to select a color close to the background color and the foreground color selected in the previous block, and the block bicolor clustering algorithm solves the small block size problem.

The multiscale bicolor clustering algorithm clusters the grid color by initializing the foreground color of the grid block to black and initializing the background color to white. As a result, pixels having a color value close to black become a foreground color and white. Pixels with near color values are extracted as the background color.

For example, if you have white text on a black background, you need to reverse it.

In this case, the inversion is performed through the foreground / background inverting algorithm.

The foreground / background inversion algorithm extracts an object, which is a collection of one foreground or background pixels, from the extracted foreground area, and uses the area occupied in the entire image and the background relation between the objects. Determine if you want to invert and perform the inversion.

Based on the K-average color clustering algorithm, the document segmentation module 100 consisting of a multiscale bicolor clustering algorithm and a foreground / background inversion algorithm receives .bmp, .jpg, and .tiff file images as input. The image is divided to generate a foreground image corresponding to a character, a background image corresponding to the image, and a bitmap image.

Commonly used color models include RGB, CMY, YUV, and YIQ color models.

For example, the RGB model is adopted for hardware or general purpose color video cameras, the CMY (Cyan, Magenta, Blue) model for color printers, and the YIQ model for color printers.

Luminance signals and chrominance signals, such as the YUV color model, can be easily compressed as compared to the RGB color model.

Since the human eye is more sensitive to the luminance signal than the color difference signal, it is not necessary to compress the color difference signal value and the luminance signal value with the same specific gravity.

For this reason, the present invention performs color difference conversion for converting an RGB color model into a YUV color model for the background image and the foreground image, respectively.

Equation 1 below converts an RGB signal to YUV.

After converting the RGB color to the YUV color by Equation 1, Y: U: V performs downsampling at 4: 1: 1 on the Y, U, and V components.

Here, downsampling means to reduce the number of image samples, and to reduce the time required to encode an image by reducing the number of samples to reduce the size of the entire image.

In this way, the foreground image and the background image generated through the document dividing module 100 perform color difference conversion, downsampling, and wavelet transform through the FDWT units 120 and 132.

The FDWT unit (Forward Discrete Wavelet Transform) (120,132) performs a wavelet transform of the lifting scheme (Lifting scheme).

The lifting scheme used to implement the wavelet transform is a recently proposed scheme.

The wavelet lifting method is a reversible method that can simply obtain an inverse wavelet transform from a forward wavelet transform, is an integer transform in which wavelet coefficients can be integers, and performs wavelet transforms. The advantage of not using a separate memory in the process.

In the present invention, the following twelve wavelet transform equations are provided.

Where x [n] is the input signal, s [n] is the lowpass subband signal, d [n] is the highpass subband signal, and s ₀ [n] is x [2n ], d ₀ [n] represents x [2n + 1].

(5/3) Convert

Forward conversion

Reverse conversion

(2/6) Convert

Forward conversion

Reverse conversion

(SP + B) Convert

Forward conversion

Reverse conversion

(9 / 7-M) Convert

Forward conversion

Reverse conversion

(2/10) Convert

Forward conversion

Reverse conversion

(5 / 11-C) Convert

Forward conversion

Reverse conversion

(5 / 11-A) Convert

Forward conversion

Reverse conversion

(6/14) Convert

Forward conversion

Reverse conversion

SPC + C Conversion

Forward conversion

Reverse conversion

(13 / 7-T) Convert

Forward conversion

Reverse conversion

(13 / 7-C) Convert

Forward conversion

Reverse conversion

(9 / 7-F) Convert

Forward conversion

Reverse conversion

The core data structure for performing the wavelet transformation is Enc_t, which is a tree structure for wavelet transformation and a structure containing all data for transformation and actual data for each component of the image. .

A tree-structured filter bank (TSFB) showing the tree structure generates a tree structure according to a decomposition level of an input image.

For example, if the decomposition level is 3, the tree structure of FIG. 3 is generated.

In the tree structure, each node includes information on a band of a corresponding level and includes all information about a buffer point and a size of an image.

After the Enc_t structure is set, a TSFB analysis process is performed to perform a wavelet lifting algorithm based on the data of the structure.

In FIG. 4, a split is performed on the actual image data.

Even part is divided into upper part, odd part is divided into lower part, upper part goes through lowpass filter through LIFT1, and lower part goes through highpass (LIFT0). highpass) filter.

By repeating this process by the number (number_of_sequence), wavelet lifting is performed for one dimension (vertical).

If lifting is performed on the horizontal axis after the one-dimensional lifting on the vertical axis, the result of the lifting on the entire image, that is, the two-dimensional image, can be obtained.

FIG. 5 illustrates a one-dimensional reversible lifting transformation process for the (5/3) transformation of Equation 2, and the transformation equation is represented by Equation 14.

FIG. 6 is a pseudo-code section for performing wavelet lifting for each pixel of one-dimensional order.

By doing this for the vertical and horizontal, respectively, four divided bands, LL, HL, LH, and HH, can be obtained.

As described above, the foreground image and the background image are converted into wavelet coefficients through color difference conversion, downsampling, and FDWT units 120 and 132, respectively.

The output wavelet coefficients are converted into gray codes again and subjected to quantization.

After the quantization process, the wavelet coefficients of the foreground image and the background image are encoded through the bit plane encoders 134 and 122 to be generated as bitstreams.

The bit plane encoders 122 and 134 start by dividing an image into a number of bit planes composed of bits.

Generate bit plane 0 around the most significant bit, and generate the least significant bit as the last bit plane.

One bit plane is divided into subblocks of appropriate sizes (2x2 to 64x64).

The bit plane encoders 122 and 134 generate bitstreams through repetition length encoding and arithmetic encoding for each subblock for the bit plane divided into subblocks.

In detail, in general, bit plane encoding is to construct a binary image by constructing a bit plane from a color image and perform compression on each binary image.

In the present invention, the binary data is generated by constructing a bit plane with respect to the wavelet coefficient value on which the wavelet transform is performed, and then performing subblock encoding.

However, a disadvantage of such bitplane coding is that small changes in coefficient values can complicate the entire bitplane.

For example, if one count value is 127 (01111111) and the adjacent count value is 128 (100000000), the count value is only 1, but all adjacent bit values appear differently in the entire bit plane.

This case is not effective for compression through redundancy of bits. Therefore, the value of each cell must be converted to graycode before performing subblock encoding.

Equation 15 below shows m gray codes g _m-1 ... For _m- bit bit planes. g ₂ g ₁ g _{0 For} conversion to ₀ .

here, Represents an XOR operation, Wow Are bit values for the wavelet coefficient.

In other words, after converting wavelet coefficient bit values to gray code, the most significant bit is gray code as it is ( The other gray code value is replaced by the XOR of the next two lower bit values ( ).

In Equation 15, 127 (01111111) is converted to a gray code value after conversion to gray code, and when a neighboring value is 128 (10000000), it is converted to gray code value after conversion to 11 million value.

That is, the change in the value between neighboring bits in the bit plane is significantly reduced.

This increases the redundancy between bits, making compression more efficient.

The wavelet coefficients converted to gray codes require at least 2 bytes of storage space, and quantization is performed to reduce the coefficients to increase the efficiency of bitplane coding and maximize compression.

In the present invention, quantization is implemented by shifting wavelet coefficient values.

The shift of the LL band, which contains the average value of the image among the subbands by wavelet transform, is relatively smaller than the LH, HL, and HH bands to minimize the distortion of the reconstructed image, and also the relative LH, HL, and HH bands. By increasing the shift, the compression ratio can be increased.

In addition, in the Y, U, and V signal values through color difference conversion, the U and V signals containing color difference values may be relatively larger than the Y signal, thereby increasing the compression ratio.

8 shows a quantization process of wavelet coefficients.

The coefficient value of the LL band representing the average value of the image is shifted by n bits, and an additional 1 bit (n + 1) is further shifted to the LH and HL bands, and an additional 2 bits (n + 2) to the HH band.

If the shift operation is performed again in the reverse direction after the shift operation is performed, as a result, the LL band has a zero value set by n bits.

After the above quantization, the wavelet coefficients are divided into subblocks on the basis of bit planes, and bit plane coding is performed on each subblock.

Through this shift quantization process, a zero value is increased in a bit of a coefficient value, and this increase of the zero value results in an increase of an "insignificant" subblock in subblock encoding.

FIG. 9A shows the bit plane of the wavelet coefficient value and FIG. 9B shows the processing of the subblock for each bit plane.

The bitplane encoders 122 and 134 check whether "significance" is present for each subblock included in the bitplane.

Importance is determined by the presence of 1 in the subblock.

If only one value exists, that subblock is considered a "significant" subblock.

If the value in the subblock is all zero, the subblock is considered to be an "insignificant" subblock and excluded from encoding, and only information about the subblock position in the entire bitplane is written to the subblock map. do.

An important subblock is encoded with the entire subblock and a value of 1 representing a significant subblock is stored in the subblock map.

The subblock encoding based on the bit plane is performed for the entire bit plane, respectively, and outputs the bitstream for the important subblocks.

At this time, the bits in the subblock are encoded while scanning from the upper left to the lower right.

An unimportant subblock does not output any bitstream, only writes the location information of the subblock in the subblock map.

Meanwhile, the JBIG encoder 110 encoding the bitmap image to generate a .jbg file may encode the binary image with high efficiency.

Here, the JBIG encoder 110 means JBIG-1 or JBIG-2.

The Joint Bi-level Image Experts Group (JBIG) encodes a bi-level image and a grayscale image.

JBIG is defined in ISO / IEC 11544 (ITU-T T.82) Recommendation for Lossless / Lossless Extrusion of Binary Images.

Since JBIG performs encoding while predicting encoded pixels from surrounding pixels, binary images can be encoded with high efficiency.

The foreground image and the background image encoded by the bit plane encoders 122 and 134 pass through the repetition length encoders 124 and 135 and the arithmetic encoders 126 and 136.

The repetition length coding method is a method of reducing a redundancy of values existing in an image by representing pixels having the same value as one code by representing a pixel block as one pixel value and the number of repetitions thereof.

For example, if the first pixel value is 234 and pixels having the same value are adjacent to each other and 14 pixels appear, the encoding is 234 @ 14.

The original image requires 15 bytes to represent a value of 15 pixels, whereas after the repetition length encoding is performed as described above, it can be represented by 2 bytes.

An adaptive arithmetic coding algorithm of arithmetic encoders 126 and 136 is used.

In the encoding method as described above, documents of .bmp, .jpg, and .tiff files containing characters and images are divided into bitmap images, foreground images of characters, and background images of images by the document splitting module 100. (Step 1)

The bitmap image is encoded through the JBIG encoder 110 and generated as a .jbg file (second step).

The foreground image is output as a wavelet coefficient through the FDWT unit 120 performing wavelet transformation using a lifting method, and is encoded through a bit plane encoder 122 that receives the wavelet coefficient as an input to generate a bitstream. (Step 3).

The background image is output as a wavelet coefficient through the FDWT unit 122 performing wavelet transform using a lifting method, and is encoded through a bit plane encoder 134 receiving the wavelet coefficient as an input to generate a bitstream ( Step 4).

The foreground image and the background image which have undergone the third and fourth processes are continuously reduced in redundancy of values existing in the image by the repetition length encoders 124 and 135, and the size of the interval is reduced by the adaptive arithmetic encoders 126 and 136. Only values that exist within this interval are encoded to complete the compression.

The decoding process is the reverse process of encoding, and the following shows the decoding process.

(1) Adaptive arithmetic decoding and repetition length decoding are performed on the bitstream of the foreground and the background.

(2) The bitstream for the bitmap is decoded by the JBIG decoder 210 to generate a bitmap.

(3) The bitstreams of the foreground and background regions are formed by performing bitplane decoding.

(4) Inverse gray code conversion is performed on the generated wavelet coefficients.

(5) Inverse wavelet transform is performed on each of the generated Y, U, and V components.

(6) Upsample the U, V component values in the generated Y, U, V components.

In this case, the upsampling ratio of the luminance value Y and the color difference values U and V is 1: 4: 4.

(7) Color difference conversion is performed from YUV color to RGB color, and color difference conversion is performed on the foreground image and the background image, respectively.

(8) The original image is reconstructed with the foreground image, the background image, and the bitmap, which are color difference converted.

The structure of the reconstruction system includes bit plane decoders 220 and 230 that receive a bitstream of a foreground image and a background image as inputs, generate wavelet coefficients, and perform inverse gray code conversion;

A JBIG decoder 210 for receiving a bitstream of a JBG file as an input and generating a bitmap image;

IDWT units (226, 236) for receiving the wavelet coefficient as an input and performing inverse wavelet transform, upsampling, and color difference conversion from YUV to RGB to generate a foreground image corresponding to a character and a background image corresponding to the image;

And a reconstruction unit 240 for reconstructing the bitmap image, the foreground image, and the background image into the original document.

The JBIG decoder 210 receives a bitstream as an input and decodes the bitstream to generate a bitmap image.

In FIG. 2, the bitstream is input and subjected to the above-described decoding process by the adaptive arithmetic decoders 220 and 230.

The repetition length decoders 222 and 232 perform decoding in a reverse process of encoding in which pixels having the same value are represented using one code.

For example, the value encoded by 234 @ 12 is decoded by indicating that the first pixel value is 234 and the same value is displayed as 14 pixels with consecutive pixels neighboring.

The bit plane decoders 224 and 234 receive the bitstreams as inputs, generate wavelet coefficients, and perform inverse gray code transformation to generate wavelet coefficients.

The IDWT units 226 and 236 receive wavelet coefficients as inputs and generate a foreground image corresponding to a character and a background image corresponding to the image.

The data structure of the Inverse Discrete Wavelet Transform (IDWT) is the same as the FDWT data structure of FIG. 3.

10 shows a synthesis process of TSFB.

The synthesis process is an inverse process of the analysis process of FIG. 4, and the original image can be obtained by performing LIFT1, LIFT0, synthesis on one dimension (vertical) by the order of repetition, and repeatedly performing the same on one dimension (horizontal). have.

By doing this for the vertical and horizontal, respectively, as shown in Figure 11 to obtain the original image for the foreground image and the background image.

In this manner, when the bitmap image, the foreground image, and the background image are generated, the image is reconstructed by the document reconstruction 240 to complete the restoration.

Reconstruction of the compressed document as described above generates a bitmap image by decoding the JBG file (first process).

Decode the encoded bitstream of the foreground image to generate wavelet coefficients, perform inverse graycode transformation, perform inverse wavelet transformation, upsample to 1: 4: 4 for YUV components, and apply YUV color To generate a foreground image by performing color difference conversion to RGB colors (second process).

Decode the encoded bitstream of the background image to generate wavelet coefficients, perform inverse graycode transformation, perform inverse wavelet transformation, perform upsampling at 1: 4: 4 on the YUV component, and apply YUV color. A background image is generated by performing color difference conversion to RGB colors (step 3).

The bitmap image, the foreground image, and the background image are reconstructed and restored to the original document (step 4).

The original experimental image used to test the performance of the document compression codec using wavelet transform and bit plane coding according to the present invention is an image having an arbitrary size.

12A (Lena image 400x255) is an image containing only an image and no text information.

FIG. 12B (Hobby image 791x565) is an image in which an image and a text are separated.

FIG. 12C (ATT image 512x512) is an image in which image and text are mixed.

The original experimental images are color images stored in RGB format.

The equation for performing wavelet transform on each image, performing bit plane coding on wavelet coefficient values, restoring the compressed image, and obtaining a peak signal-to-noise ratio (PSNR) according to the equation Same as 16

Here, 255 is a maximum value in pixels, and MSE (Mean Squared Error) is calculated by Equation 17 as the sum square of the error between the original experimental image and the reconstructed image for all pixels.

Where W is the image width and H is the height of the image.

13A and 13B show compression ratios for background and foreground images, and FIGS. 14A and 14B show PSNR values for reconstructed images.

Here, it can be seen that the average compression ratio of the background image is about 100: 9 and that of the foreground image is about 100: 7.

For background images, (5/3) and (5 / 11-C) wavelet lifting show the highest compression.

(9 / 7-F) has a relatively poor compression ratio, but the PSNR shows the highest value.

(5/3) and (5 / 11-C) wavelet lifting also show relatively high PSNR with high compression ratio.

In the foreground image, (5/3) and (2/6) wavelet lifting show the highest compression ratio.

(5/3) Wavelet lifting shows high PSNR at the same time as compression rate.

On the other hand, (9 / 7-F) can be seen to be the most inefficient for both compression rate and PSNR.

15A, 15B, and 15C show reconstructed images of the compressed files of the experimental images FIGS. 12A, 12B, and 12C.

Comparing FIG. 12 with FIG. 15, it can be seen that the experimental image and the reconstructed image have little visual difference.

As described above, according to the present invention, the text and the image may be separated and compressed and restored as the text and the video as the image.

In addition, according to the present invention, since 12 integer transforms can be performed, the integer transform can be selected more precisely according to the input image, so that the loss caused when the image is compressed and reconstructed is significantly increased. Can be reduced.

In addition, by converting the wavelet coefficients to gray code, it improves the compression effect by increasing redundancy in the bit plane, and by applying quantization at different ratios for each subband, thereby increasing the compression rate without lowering the visual quality of the image. Can be.

Claims (13)

A document splitting module for generating a document including text and an image into a foreground image corresponding to a character, a background image corresponding to a background and an image, and a bitmap image; A JBIG encoder for generating a JBG file by receiving the bitmap image as an input and encoding the bitmap image; An FDWT unit which receives the foreground image and the background image as inputs, performs color difference conversion and downsampling, and performs wavelet transform by a lifting method to output wavelet coefficients; A bit plane encoder which receives the wavelet coefficients as an input, performs gray code transformation and quantization, and generates a bit stream by performing encoding by a bit plane method; Compression system for documents containing embedded text and images. A bit plane decoder which receives the bitstreams of the foreground image and the background image as inputs, generates wavelet coefficients, and performs inverse gray code conversion; A JBIG decoder which receives a bitstream of a JBG file as an input and generates a bitmap image; An IDWT unit which receives the wavelet coefficient as an input and performs inverse wavelet transform, upsampling, and color difference conversion to generate a foreground image corresponding to a character and a background image corresponding to the image; A document reconstructing unit for restoring the bitmap image, the foreground image, and the background image to an original document; Restoration system for documents containing embedded text and images. A first step of dividing the entire image of the document including the character and the image into a bitmap image, a foreground image of the character, and a background image of the image; Generating a JBG file by encoding the bitmap image; Performing a wavelet transform on the foreground image by a lifting method to output wavelet coefficients, and encoding the wavelet coefficients to generate a bitstream; Performing a wavelet transform on the background image by a lifting method to output wavelet coefficients, and encoding the wavelet coefficients to generate a bitstream; How to compress documents that contain embedded text and images. The method of claim 3, wherein the division of the document in the first process, Cluster the entire image into the foreground and background colors, 1.The process of initializing the foreground color to black and the background color to white, 2. The process of comparing each pixel value of the image and the distance between the foreground color and the background color, and including them closer to each other; 3. The process of obtaining the average of each pixel included in the foreground and the pixel included in the background and updating them to the foreground color and the background color, 4. Repeat steps 2 and 3 until each color is collected, A method of compressing a document containing text and images, characterized in that it takes place through. The method according to claim 3 or 4, wherein color difference conversion for converting an RGB color model into a YUV color model is performed on the background image and the foreground image divided in the first step, and 4: 1 for Y, U, and V components. A method of compressing a document containing text and images, characterized by performing downsampling at: 1. The method according to claim 3, wherein the wavelet coefficient value is converted into a gray code, and the coefficients converted into the gray code are subjected to different quantization for each band (LL, LH, HL, HH), and then encoding is performed. How to compress documents containing text and images. The method of claim 6, wherein the quantization is implemented by shifting coefficient values. The method of claim 7, wherein the shift of the LL band containing the average value of the image in the band is relatively less than the LH, HL, HH band. The method of claim 3 or 6, wherein the encoding divides wavelet coefficients into subblocks on a bit plane basis, and performs bit plane encoding on each subblock. 10. The method of claim 9, wherein whether the subblock is a significant subblock containing a value of 1 by examining whether or not the significance is considered according to the presence or absence of 1 present in the subblock. If the bitstream is generated by encoding and the subblock is an insignificant subblock having only 0, only the information about the subblock position in the entire bit plane is stored in the subblock map except for encoding. How to compress documents that contain text and images. 11. The method of claim 10, wherein no bitstream is output when the subblock is a non-significant subblock consisting only of zeros. The method of claim 3, wherein the third and fourth wavelet lifting is performed. Splitting the actual image data, divide the even part into the upper part and the odd part into the lower part. The upper part goes through a lowpass filter, the lower part goes through a highpass filter, A method of compressing a document containing characters and images, characterized in that the sequence is performed repeatedly. Generating a bitmap image by decoding the encoded JBG file; Decode the encoded bitstream of the foreground image to generate wavelet coefficients, perform inverse graycode transformation, perform inverse wavelet transformation, upsample to 1: 4: 4 for YUV components, and apply YUV color Performing a color difference conversion to RGB colors to generate a foreground image; Decode the encoded bitstream of the background image to generate wavelet coefficients, perform inverse graycode transformation, perform inverse wavelet transformation, perform upsampling at 1: 4: 4 on the YUV component, and apply YUV color. Generating a background image by performing color difference conversion to RGB colors; A fourth step of reconstructing and reconstructing the bitmap image, the foreground image, and the background image into an original document; How to restore documents that contain embedded text and images.