Classifications

• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
  • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
  • H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
  • H04N19/134—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or criterion affecting or controlling the adaptive coding
  • H04N19/136—Incoming video signal characteristics or properties
• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
  • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
  • H04N19/60—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using transform coding

Definitions

• the present invention relates to a two-dimensional or three-dimensional image information compression method, and more particularly, to an image information compression method for compressing image information using a feature separation code of an image.
• image separation is performed by separating image information for each pattern included in the image information into two pieces of feature data, and encoding each piece of feature data to compress the image information.
• image information compression methods for example, Jonson K. Yan, David J. 3 ⁇ 4 akrlson, Jin coding of Image Basedona Two—Component Source Model, IEE Transacti -25, No. 11 (1977), according to which the original image data is first made one-dimensional by raster scanning.
• the edge component and contrast component of the image are captured in the discontinuous component d (X, y), and the texture component of the image is captured in the remaining component r (x, y). They are separated and the optimal coding is performed for each feature component.
• the spectrum of the raster-scanned one-dimensional original data contains many high-frequency components. Therefore, smooth one-dimensional original data In this case, the edge component and the contrast component are not sufficiently taken into the smoothing function and remain in the remaining components. Therefore, there was a problem that sufficient coding efficiency could not be obtained, and that if the information of the remaining components was reduced in order to increase the coding efficiency, image degradation would increase.
• a problem to be solved by the present invention is to provide a powerful image information feature separation and encoding method, which can effectively separate an image for each feature, and has a high encoding efficiency and a small image quality degradation. Is to provide. Disclosure of the invention
• an image information compression method is directed to an image information compression method for performing feature separation encoding of an original image, wherein the two-dimensional or three-dimensional original image is one-dimensionally simulated by pseudo-Hilbert scanning.
• the first feature information is extracted by converting the data into original data and approximating the one-dimensional original data with a predetermined function having a greater redundancy than the one-dimensional original data to generate first feature data.
• the feature information is encoded into first feature code data
• the second feature information is extracted by subtracting the first feature information from the original image
• the second feature information is encoded to obtain second feature code data.
• a two-dimensional or three-dimensional original image is converted into one-dimensional original data having spectral characteristics having a small number of high frequency components and a large number of low frequency components by pseudo-Hilbert scanning.
• the pseudo-Hilbert scanning has a characteristic of scanning an image for each square or rectangular block.
• an image has a characteristic that the correlation between pixel values in a block is large and the change in pixel value in a block is moderate on average. Therefore, the power spectrum of the one-dimensional original data obtained by the pseudo-Hino level scan is concentrated in the low-frequency region, and has a characteristic that the high-frequency component is extremely small as compared with the commonly used raster scan.
• This one-dimensional original data is approximated by a predetermined function with greater redundancy than the one-dimensional original data. Then, this approximate function value is used as the first feature data.
• the second feature data is generated by subtracting the first feature data from the original image.
• the spectral characteristics of the one-dimensional original data have few high-frequency components. Therefore, by appropriately selecting a predetermined function, the approximation error when approximating with the predetermined function can be reduced. Can be made smaller than the absolute value average of each value of the one-dimensional original data, and the characteristic components of the original image can be sufficiently separated.
• the contrast component and edge component of the original image are mainly captured in the first feature data, and the texture component is mainly captured in the second feature data.
• the first feature data has the same code amount as the original image, but the redundancy is increased compared to the original image, the first feature data can be encoded with high efficiency.
• the second feature data is obtained by subtracting the first feature data from the original image
• the second feature data mostly includes the contrast component and the edge component of the original image, and mainly includes only the high frequency component of the original image. Therefore, since the spectral characteristic of the second feature data is biased toward the high frequency side, by adopting a code method suitable for this, the second feature data can be encoded with high coding efficiency. It is possible to Therefore, the entire original image can be encoded with high encoding efficiency.
• the “original image” is not limited to a still image, but may be a moving image composed of a plurality of temporally continuous frame sequences.
• the term “two-dimensional or three-dimensional” means that in the case of a still image, the original image is two-dimensional.
• Feature separation encoding refers to an image encoding method in which an image is separated for each feature component included in the image and encoded for each feature component.
• Pseudo-Hilbert scanning refers to scanning performed along a Hilbert curve extended so as to be applicable to a rectangular area or a cubic area. For details, refer to JP-A-11-353453, J-D-II, No. 10, pp. 2864-2867, Oct. 1997, S.
• the “pseudo-Hilbert scan” is not limited to two-dimensional pseudo-Hilbert scanning, but may be three-dimensional pseudo-Hilbert scanning.
• the original image has only one frame, so a two-dimensional pseudo-Hilbert scan is used.However, in the case of a moving image, the original image is a three-dimensional image.
• the two-dimensional pseudo-Hilbert run is disclosed in IEICE (D-11), Vol. J 80-D-II, No. 10, pp. 2864-2867, Oct. 1997, etc.
• the three-dimensional pseudo-Hilbert scanning is described in JP-A-11-353453, IEICE (D-I1), Vo 1. J81-D-II, No. 10, p. p. 2483-2486, Oct. 1998, and JP-A-2000-253396.
• the “predetermined fraction” does not particularly limit the function form, and for example, a step function, a polygonal function, a spline function, or the like is used. However, it is preferable to use the step function in consideration of the high speed of the calculation process and the good edge component capturing characteristics. Also, “large in redundancy” means that the autocorrelation coefficient of the first feature data is larger than that of the one-dimensional original data.
• first feature data can be encoded with high encoding efficiency.
• First feature data refers to one-dimensional original data This refers to data consisting of an approximated one-dimensional series.
• first feature information refers to information possessed by the first feature data.
• Second feature information refers to information obtained by subtracting the first feature information from the original image, and information having a data sequence obtained by subtracting the first feature data from the one-dimensional original data; and It includes information of an image obtained by subtracting the first feature image from the original image in which one feature data is filled in a two-dimensional or three-dimensional space by pseudo-Hilbert scanning.
• the method of “encoding the first feature information” or “encoding the second feature information” is not particularly limited, and encoding methods such as general quantization and edge coding are used. Used.
• the image information compression apparatus of the present invention is an image information compression apparatus for performing feature separation coding of an original image, wherein the pseudo-Hilbert scanning for converting the two-dimensional or three-dimensional original image into one-dimensional original data by pseudo-Hilbert scanning.
• An information extraction unit, and a second encoding unit that encodes the second characteristic information to generate second characteristic code data.
• the entire original image can be encoded with high encoding efficiency.
• the original image is composed of a plurality of frame sequences that are continuous in the time direction
• a predetermined frame has greater redundancy than the one-dimensional original data.
• the first feature information is extracted by approximating the one-dimensional original data by a predetermined function to generate first feature data
• the first feature data is stored in first feature data storage means.
• the first difference data is generated by subtracting the first feature data of the past frame stored in the first feature data storage means from the one-dimensional original data.
• the first feature data can be generated by approximating the first difference data with a predetermined function having a large redundancy.
• the correlation between successive frames may be large.
• a component of the original image that has a small change in the time direction in the original image is obtained. Usually removed. As a result, the coding efficiency of the original moving image can be further improved.
• the “predetermined frame” refers to a frame that has no correlation or a small correlation with a past frame, such as the first frame or a frame immediately after a change when a scene change is performed.
• “Past frame” refers to a frame in which image information compression has been performed before a frame in which image information compression is currently performed.
• second feature data is generated by subtracting the first feature data from the one-dimensional original data, and the second feature data is subjected to a pseudo-Hilbert run.
• the second feature image is restored by filling in a three-dimensional or three-dimensional space, and the second feature image is expanded into a space spanned by the base vector using the base vector stored in the base vector table.
• the component values of the second feature image with respect to a predetermined base vector in a space spanned by the base vector can be code-equipped as second feature code data.
• the first feature image is restored by filling the first feature data into a two-dimensional or three-dimensional space by pseudo-Hilbert scanning, and the first feature image is obtained from the original image.
• the second feature image is generated by using the base vector stored in the base vector table, and the second feature image is developed into a space spanned by the base vector.
• the component values of the second feature image with respect to a predetermined base vector in the space to be extended can be encoded as second feature code data.
• the one-dimensional original data has spectral characteristics with few high-frequency components and many low-frequency components.
• the contrast components DC components and low-frequency components
• edge components of the original image are obtained.
• the outline component of the image is mainly taken in the first feature information, and the remaining texture components are mainly taken in the second feature information.
• This texture component has a characteristic that the power spectrum is shifted to the high frequency side. Therefore, when the second feature information is expanded into a space spanned by base vectors, the spectrum tends to concentrate on some base components corresponding to the medium-high frequency components other than the DC component and the low-frequency components.
• the second feature image is identified.
• the basis vectors for expanding to these basis components are obtained, and these basis vectors are stored in the basis vector table. Then, when the base vector stored in the base vector table is expanded into a space where each second feature image is spanned by the base vector, the base vector is expanded into a valid base component even with each second feature image. It is regarded as a possible basis vector.
• the second feature image is expanded into a space spanned by the base vector using the base vector stored in the base vector table in advance (that is, the second feature image and the base vector).
• the “space spanned by basis vectors” refers to a space in which patterns are mapped by feature extraction, and this mapping includes Karhunen-Loeve transform (hereinafter referred to as “KL transform”), discrete cosine.
• Orthogonal transforms such as transform (hereinafter referred to as “DCT”) and discrete sine transform (hereinafter referred to as “DST”) are used.
• “Expanding in the space spanned by the base vector” means mapping the second feature image to the space spanned by the base vector, and more specifically, the mapping between each base vector and the second feature image. It is performed by inner product calculation.
• the KL transform when each pixel of the second feature image is mapped to the feature space, how many variances of the basis components of each pixel are large;
• DTC when used for feature extraction, it refers to the basis vector selected in the order of frequency components with the largest spectrum of the second feature image.
• the second feature image When developing the second feature image into the space spanned by the base vectors, it is not always necessary to develop the second feature image for all base vectors. Only the second feature image may be developed, and the component values for the other basis vectors may be regarded as 0. In other words, using the test pattern image, several base vectors whose component values have significant values on average are selected in advance, and the second feature image is developed only for those base vectors. It can be done. Even in this case, the amount of decrease in the SZN ratio of the image quality is small, and the amount of calculation when the second feature image is expanded is reduced, so that high-speed processing can be performed while suppressing the increase in the amount of coding noise. Become.
• second feature data is generated by subtracting the first feature data from the one-dimensional original data, and the second feature data is generated using a base vector stored in a base vector table.
• the second feature data is expanded in a space spanned by the base vector, and the component values of the second feature data for a predetermined base vector in the space spanned by the base vector are encoded to encode the second feature data. It can be used as feature code data.
• the SZN ratio of the image quality is not inferior, and high-speed image encoding can be performed.
• the error data is orthogonally transformed and compressed as it is in one dimension, so that the amount of calculation is small and faster image encoding is possible.
• the component value of the second feature image or the second feature data for the base vector is encoded. Into the second feature code data.
• the pseudo-Hilbert scan is used when the original image is made one-dimensional, most of the contrast component and the edge component of the original image are taken into the first feature code data. Therefore, of the component values of the second feature image for each base vector in the space spanned by the base vector, only the component values whose component values are equal to or greater than a predetermined threshold are left, and the remaining component values whose component values are small Even if is regarded as 0, there is little deterioration in image quality when the original image is visually restored. This is because when humans perceive an image, the features of the image are recognized from the outline and the global pattern, so if the contrast component and the edge component are sufficiently reproduced, the reproducibility of fine textures is somewhat poor.
• the “predetermined threshold value” can be freely set according to the type and purpose of the image to be coded, and if a large threshold value is set, the coding efficiency is improved, but the coding is improved.
• the S / N ratio of the resulting image decreases.
• the base feature to expand the second feature image or the second feature data is based on a cumulative error between the one-dimensional original data and the first feature data. Determine the vector, expand the second feature image or the second feature data with respect to the determined base vector, and calculate the component value of the second feature image or the second feature data with respect to the base vector. It can be encoded and used as second feature code data.
• a set of base vectors in a space spanned by base vectors whose component values are significant is appropriately set according to the pattern of the original image, and the component values are set only for the set of base vectors. Is calculated, and the component values corresponding to the other base vectors are regarded as 0. Encoding efficiency can be increased without increasing coding noise. Also, since it is not necessary to calculate the component values of the second feature image for all the base vectors in the space spanned by the base vectors, high-speed coding can be performed.
• the “cumulative error” refers to a value obtained by adding the absolute value or the square value of the difference between the one-dimensional original data and the first feature data for each pixel for all pixels.
• a relation between the magnitude of the cumulative error and the component value for each base vector is determined in advance using a test pattern, and the cumulative error is determined.
• the correspondence between the set of base vectors whose component values are significant with respect to the size is registered in the judgment table, and when compressing image information, refer to the judgment table according to the magnitude of the accumulated error. Then, a method of determining the base vector is adopted.
• the first feature data is composed of a data sequence according to a step function of changing in a stepwise manner at predetermined intervals, and the data value of the first feature data in each interval is the data value of the one-dimensional original data.
• the data series is divided into the sections so that the accumulated error over the entire section between the one-dimensional original data and the first feature data is minimized, equal to the average value in the section. It can be.
• the algorithm of the approximation calculation process is simple, the amount of calculation required for the approximation calculation process is small, and the speed can be increased. Also, by approximating with a step function divided into each section so that the accumulated error between the one-dimensional original data and the first feature data is minimized, sharp edges in the image such as mosquito noise are smoothed.
• the outline portion of the image is not blurred. Therefore, in the first feature code data, the outline portion in the image is accurately stored, and when the first feature image is restored from the first feature code data, the outline of the image in the first feature image is clear. Yes, it is possible to accurately recognize the contents of images visually.
• the first feature data is composed of a data sequence according to a step function that changes stepwise in a predetermined section, and the data value of the first feature data in each section. Is equal to the average value of the one-dimensional original data in the section, and the data series is such that the cumulative error in each section between the one-dimensional original data and the first feature data is equal to or less than a predetermined threshold.
• the section can be divided into the sections.
• FIG. 1 is a block diagram of an image information compression device and an image restoration device according to Embodiment 1 of the present invention.
• FIG. 2 is a flowchart illustrating an image information compression method according to Embodiment 1 of the present invention.
• FIG. 3 is a diagram comparing the power spectrum of one-dimensional original data obtained by one-dimensionalization of a certain image by raster scanning and one-dimensional original data obtained by one-dimensionalization by pseudo-Hilbert scanning.
• FIG. 4 is a diagram showing the relationship between the one-dimensional original data and the first feature data.
• FIG. 5 is a flowchart illustrating a method of restoring an original image from hierarchical data encoded by the image information compression method according to the first embodiment.
• FIG. 6 is a block diagram of an image information compression device and an image decompression device according to Embodiment 2 of the present invention.
• FIG. 7 is a block diagram of an image information compression device and an image restoration device according to Embodiment 3 of the present invention.
• FIG. 8 is a flowchart showing an image information compression method according to Embodiment 3 of the present invention.
• FIG. 9 is a flowchart illustrating a method of restoring an original image from hierarchical data encoded by the image information compression method according to the third embodiment.
• FIG. 10 is a block diagram of an image information compression device and an image restoration device according to Embodiment 4 of the present invention.
• FIG. 11 is a flowchart illustrating an image information compression method according to Embodiment 4 of the present invention.
• FIG. 12 is a flowchart illustrating a method of restoring an original image from hierarchical data encoded by the image information compression method according to the fourth embodiment.
• FIG. 1 is a block diagram of an image information compression device and an image decompression device according to Embodiment 1 of the present invention.
• 1 is an image compression device
• 2 is an image restoration device
• the image compression apparatus 1 includes an image input unit 3, a frame memory 4, a pseudo-Hilbert scanning unit 5, a first feature information extracting unit 6, a first code fetching unit 7, a second feature information extracting unit 8, and a second encoding unit. Means 9 are provided.
• the image input unit 3 has a function of inputting a two-dimensional or three-dimensional original image from outside the device, and is specifically configured by an image input device such as a scanner or a video camera.
• the frame memory 4 stores the original image input to the image input unit 3. Is temporarily stored.
• the pseudo-Hilbert scanning means 5 converts the original image stored in the frame memory 4 into one-dimensional original data by pseudo-Hilbert scanning.
• the first feature information extracting means 6 generates the first feature data by approximating the one-dimensional original data with a predetermined function, and stores the accumulated error in the accumulated error memory 10.
• the first code selector 7 generates first characteristic code data by compressing information of the first characteristic data by entropy coding.
• the second feature information extracting means 8 includes a cumulative error memory 10, a second feature data generating means 11, a second feature image restoring means 12, a judging means 13, a judging table 14, a base vector storage means 15 And orthogonal transformation means 16.
• the cumulative error memory 10 stores the cumulative square error between the one-dimensional original data and the first feature data.
• the second feature data generation means 11 generates one-dimensional second feature data by calculating a difference between the one-dimensional original data and the first feature data.
• the second feature image restoring means 12 restores the one-dimensional second feature data into a two-dimensional or three-dimensional second feature image by pseudo-Hilbert scanning.
• the judging means 13 refers to the judgment table 14 based on the cumulative square error stored in the cumulative error memory 10 to determine a set of base vectors for which the orthogonal transform means should calculate the inner product. .
• the determination table 14 stores a list indicating the correspondence between the component vectors and the set of base vectors having a significant magnitude in accordance with the magnitude of the cumulative square error. This correspondence is determined in advance by using a test pattern to determine the relationship between the magnitude of the accumulated error and the component value for each basis vector, thereby obtaining the basis vector required for the magnitude of the accumulated squared error. Identify and register a set.
• the basis vector storage means 15 stores a basis vector table including a set of basis vectors when orthogonally transforming the second feature image.
• a basis vector table including a set of basis vectors when orthogonally transforming the second feature image.
• the base vectors for developing the feature images into their base components are obtained, and these base vectors are stored in the base vector table.
• the orthogonal transformation means 16 is a base vector storage means. The inner product of each base vector stored in 15 and the second feature image is calculated.
• the second encoding unit 9 has a function of encoding the set of inner products to generate second feature code data, and includes a quantization unit 17 and an entropy encoding unit 18.
• the image restoration unit ft2 includes a first decoding unit 19, a first feature image restoration unit 20, a second decryption unit 21, a second feature image restoration unit 22, an original image restoration unit 23 , An image memory 24, and output means 25.
• the first decoding unit 19 decodes the first feature code data and restores the first feature data.
• the first feature image restoring means 20 restores the first feature data into a two-dimensional or three-dimensional first feature image by pseudo-Hilbert scanning.
• the second decryption means 21 decodes the second feature code data and restores a combination of inner products and a base vector.
• the second feature image restoring means 22 restores a two-dimensional or three-dimensional second feature image by applying an inverse matrix of a base vector to a set of inner products.
• the original image restoring means 23 restores the original image by adding the second feature image to the first feature image.
• the image memory 24 stores the restored original image.
• the output unit 25 is a device that outputs a restored image, and includes a display, a printer, a facsimile, and the like.
• FIG. 2 is a flowchart illustrating an image information compression method according to Embodiment 1 of the present invention.
• the original image G from the image input unit 3. Is input, the original image G. Is stored in frame memory 4 (S l). In this case, if the input original image is a still image, the original image G. Is stored as a two-dimensional original image, and if it is a moving image, the original image G. Is stored as a three-dimensional original image in which many two-dimensional image frames overlap in the time axis direction.
• the pseudo-Hilbert scanning means 5 performs pseudo-Hilbert scanning on the original image G 0 stored in the frame memory 4 to obtain one-dimensional original data ⁇ g J: i
• S is the original image G.
• S include Indicates the total number of pixels.
• the one-dimensional original data converted one-dimensionally by the pseudo-Hilbert scanning is the original image G. It has the characteristic that the spectrum has very few high-frequency components compared to the one converted into one-dimensional original data by raster scanning or zigzag scanning.
• Figure 3 is a diagram comparing the power spectrum of one-dimensional original data obtained by one-dimensional rasterization of the same image and one-dimensional original data obtained by one-dimensional pseudo-Hilbert scanning. Is the power spectrum of the one-dimensional original data that has been made one-dimensional by raster scanning, and (b) is the power spectrum of the one-dimensional original data that has been made one-dimensional by pseudo-Hilbert scanning.
• the image is made one-dimensional by repeating the operation of scanning to the right end of the image, returning to the left end of the image, and scanning again.
• the pixel value changes from the right end to the left end of the image, a sharp change occurs in the pixel value, so that a high frequency component is generated, and the power spectrum spreads over a considerably high frequency band.
• scanning is performed for each block of an image along a pseudo-Hilbert curve, which is a kind of space filling curve.
• an image has a characteristic that the correlation between pixel values in a block is large and the change in pixel value in a block is moderate on average.
• the power spectrum of the one-dimensional original data obtained by pseudo-Hilbert scanning is concentrated in the low-frequency region, and the high-frequency components are extremely small compared to the commonly used raster scan.
• the one-dimensional original data that has been made one-dimensional by the pseudo-Hilbert scan has few components that change rapidly and many components that change slowly. This means that the error can be reduced even when the one-dimensional original data is approximated by an appropriate function with large redundancy. Therefore, by selecting a function suitable for extracting a specific feature of an image as an approximation function, it becomes possible to separate only a specific feature component of the image very efficiently.
• the first feature data is set so that the cumulative square error e (p k , 1 k ) between the one-dimensional original data 0 and the first feature data in each section k is equal to or less than a predetermined threshold ⁇ .
• section division is performed by the following sequential processing.
• the dashed line in FIG. 4A is the one-dimensional original data, and the solid line is the first feature data.
• Fig. 4 (b) is the first feature data, and the scale of the vertical axis is enlarged from Fig. 4 (a).
• the first feature data changes abruptly in a step-like manner at the edge part (A, B, C, etc. in Fig. 4) of the one-dimensional original data (original image).
• the one-dimensional original data is averaged and changes smoothly, so that the edges and the contrast of the original image are well preserved, and the texture components other than the edges and the contrast are removed.
• the data sequence obtained by approximating the one-dimensional original data G x by the step function divided into each section so that is minimized may be used as the first feature data.
• the encoding means 7 performs entropy encoding of the first feature data (S4), and outputs this as first feature code data (S5).
• the first feature code data includes marker codes such as an image start code (SOF), a definition parameter group (DP), an image end code (EO I) used for encoding by the first code selector. Will be added. This is necessary when restoring the first characteristic code data.
• the orthogonal transformation means 16 converts the second feature image G 2 into M 0 rows XN.
• p Ri subscript der representing the position of the partial second feature image in the row direction in the second feature image G 2
• q is the column direction of the partial second feature image in the second feature image G 2
• I is a subscript representing the position of the pixel in the row direction in the partial second feature image f p , q
• j is the subscript representing the pixel in the column direction in the partial second feature image i p , q . This is a subscript indicating the position.
• the second feature image G 2 is divided into L in the row direction and into C in the column direction.
• the base base vector ⁇ a k (i,; i ) ⁇ is a basis vector in the case of orthogonal transform part second feature image ⁇ f p, q (i, j) ⁇ , as the orthogonal transformation, KL transform, DCT, DST, etc. are used.
• the inner products b p , q and k are Where f p , q and a k are M.
• XN which is the next vector.
• the inner product force S, partial second feature image in the space spanned by the basis base vector becomes coordinate value of ⁇ f p, g (i, j) ⁇ .
• the orthogonal transformation means 16 does not calculate the inner product of all the basis vectors a k (i, j), and the eigencomponents of the partial second feature image f p , q (i, j) are significant values.
• the dot product is output only for the base vector a k (i,; j).
• & 1 (i, j), a 2 (i,]), ⁇ , a m in the order of basis components that can accurately represent the pattern of the partial second feature image f p , resort(i, j)
• the base vectors are arranged as (i, j)
• the number of pairs of basis vectors used for inner product calculation Hi is determined by referring to the decision table based on the value of the matrix vector. ⁇ ⁇ a x (i, j) , a 2 (i, j), ⁇ ⁇ ⁇ , a m (.
• the orthogonal transformation means 16 does not necessarily need to perform M 0 XN 0 inner product operations. Instead, m inner product operations need only be performed on the effective feature axis, so that the amount of calculation is reduced, and the speed of information compression processing can be increased.
• the orthogonal transformation means 16 performs an inner product operation on all the sets of base vectors ⁇ ak (i, j) ⁇ , and when the inner product is a predetermined threshold / It is also possible to output only three or more items and output 0 for other items.
• the second encoding means 13 encodes the set of inner products ⁇ b p , q , k ⁇ (S10), Output as the second feature code data (S11).
• the code E is quantized by the quantization unit 17 and then entropy-encoded by the entropy encoding unit 18 to perform information compression.
• the second feature code data includes the image start code (SOF), the definition parameter group (DP) used for coding in the first code selector, and the image end code (EOI). Code is also included.
• the partial second feature image is orthogonally transformed, and only the inner product with respect to the feature axis is encoded. Therefore, high code efficiency can be obtained.
• FIG. 5 is a flowchart illustrating a method of restoring an original image from hierarchical data encoded by the image information compression method according to the first embodiment.
• the first decoding means 19 executes the code-encoded image included in the first feature code data. Is decrypted and the first feature data is restored (S21).
• the first feature image restoring means 20 restores a two-dimensional or three-dimensional first feature image G 0 , by filling the first feature data G x , into a two-dimensional or three-dimensional space by pseudo-Hilbert scanning. (S22).
• the second decoding means 21 receives the second feature code data transmitted from the second coding means 9 (S23), the second decoding means 21 decodes the coded image included in the second feature code data.
• the decryption is performed to restore the inner product set ⁇ b p , q , k ⁇ (S24).
• the second feature image restoration hand stage 22 portions the second feature image ⁇ f p, q '(i , j) ⁇ Restore, restored partially the second feature image ⁇ ⁇ , q, (i , j) ⁇ from the second feature image G 2 , (S25) 0
• the inner product set ⁇ b p , q> k ⁇ is represented by equation (5)
• the partial second feature image ⁇ f p , q '(i, j) ⁇ is restored by performing the following operation. (6)
• the M 0 XN 0- order betatonore a k is the inverse vector of the base vector a k . Since the ineffective components of the inner product set ⁇ b p , q , k ⁇ are approximated by 0, the restored second feature image G 2 ′ is different from the original second feature image G 2. Does not exactly match, but gives a good approximation.
• the original image restoring means 23 restores the original image G 0 ′′ by adding the first feature image G 0 ′ and the restored second feature image G 2 , (S 26). . Are stored in the image memory 24 and output to the output means 25 such as a display (S27).
• FIG. 6 is a block diagram of an image information compression device and an image restoration device according to Embodiment 2 of the present invention.
• 1 is an image compression device
• 2 is an image decompression device
• 3 is an image input unit
• 4 is a frame memory
• 5 is a pseudo-Hilbert running means
• 6 is a first feature information extracting means
• 7 is a first feature information.
• Encoding means 9 is second encoding means, 10 is cumulative error memory, 13 is judgment means, 14 is judgment table, 15 is base vector storage means, 16 is orthogonal transformation means, 17 is quantization unit, 18 Is an entropy encoding unit, 19 is a first decoding unit, 20 is a first feature image restoration unit, 21 is a second decoding unit, 22 is a second feature image restoration unit, 23 is an original image restoration unit, and 24 is an original image restoration unit.
• the image memory 25 is an output means, which are the same as those in FIG.
• the present embodiment is characterized in that the second feature information extracting unit 30 includes a first feature image restoring unit 31 and a second feature image generating unit 32.
• the first feature image restoring means 31 fills the two-dimensional or three-dimensional space with the first feature data G ,, by pseudo-Hilbert scanning, Restore to the first feature image G 0 '.
• the second feature image G 2 obtained by it this, the orthogonal transform unit 16 and by Ri compressed second coding means 9, which is constituted in such a manner as to encode.
• FIG. 7 is a block diagram of an image information compression device and an image restoration device according to Embodiment 3 of the present invention.
• 1 is an image compression device
• 2 is an image decompression device
• 3 is an image input unit
• 4 is a frame memory
• 5 is a pseudo-Hilbert scanning means
• 6 is first feature information extraction means
• 7 is a symbol.
• Coding means 9 is second coding means
• 10 is cumulative error memory
• 11 is second feature data generation means
• 13 is judgment means
• 14 is judgment table
• 15 is base vector storage means
• 17 is base vector storage means.
• Quantization unit 18 is an entropy encoding unit
• 19 is a first decoding unit
• 20 is a second encoding unit
• 24 is an image memory
• 25 is an output unit. The same reference numerals are given and the description is omitted.
• the second feature information extracting means 33 encodes the component value of the second feature data with respect to a predetermined base vector in a space spanned by the base vector, and generates the second feature code data. Characterized in that the orthogonal transform means 34 is provided. Further, the image restoration device 2 includes a second feature data restoration unit 35, a one-dimensional original data restoration unit 36, and an original image restoration unit 37.
• FIG. 8 is a flowchart showing an image information compression method according to Embodiment 3 of the present invention.
• S31 to S36 are the same as the operations of S1 to S6 in FIG. 8.
• the orthogonal transformation means 34 converts the one-dimensional second feature data G 2 ′ to M.
• ⁇ As the value of, 64 pixels are used.
• the base vector storage means 15 stores the one-dimensional second feature data in an orthogonal A base vector table including a set of base vectors at the time of conversion is stored.
• the inner product b p , k of the vector a k (i, j) and the partial one-dimensional second feature data ⁇ ⁇ (i) is calculated (S38). That is, in the present embodiment, the inner products b p and k are
• the orthogonal transformation means 34 does not output the inner product of all the basis vectors a k (i), and outputs the basis where the eigen component of the partial one-dimensional second feature data f p (i) becomes a significant value.
• the basis vector ⁇ a k (i) ⁇ is obtained by using a plurality of test patterns to calculate the basis vector corresponding to the most effective feature axis in the orthogonal transformation of the partial one-dimensional second feature data ⁇ p (i).
• the base order such as a (i), a 2 (i), ⁇ , a m (i)
• the number m of pairs of base vectors to be used is determined.
• the orthogonal transformation means 34 calculates the inner product value only for the set of base vectors ⁇ ai (i), a 2 (i), ⁇ - ⁇ , a m (i) ⁇ of the number determined by the determination means 13. Output, and output 0 for other basis vectors.
• the orthogonal transform means 16 is not necessarily M. It is not necessary to perform the inner product operation twice, and it is sufficient to perform the inner product operation m times for the effective base vector. Therefore, the amount of calculation is reduced, and the information compression process can be performed at high speed.
• the two-dimensional or three-dimensional second Without restoring the image, the inner product of the one-dimensional second feature data G 2 , with the basis vector ⁇ ak (i) ⁇ is calculated, compressed, and coded. It is suitable in the case where is required.
• FIG. 9 is a flowchart illustrating a method of restoring an original image from hierarchical data encoded by the image information compression method according to the third embodiment.
• the first decoding means 19 receives the first feature code data transmitted from the first coding means 7 (S41), the coded data included in the first feature code data is received. The image is decoded, and the first feature data Gi 'is restored (S42).
• the second decoding means 21 when the second decoding means 21 receives the second feature code data transmitted from the second coding means 9 (S43), the second decoding means 21 decodes the code included in the second feature code data.
• the decoded image is decoded to recover the inner product set ⁇ b p , k ⁇ (S 21).
• the one-dimensional second feature data restoring means 35 restores the second feature data of a portion one-dimensional ⁇ f (i) ⁇ , the restored portion one-dimensional second feature data ⁇ f p, (i) ⁇ , The one-dimensional second feature data G 3 , is restored (S45).
• the restored one-dimensional second feature data G 3 ′ is the original one-dimensional second feature data. Although not completely coincide and G 3, it gives a good approximation.
• the one-dimensional original data restoring means 36 restores the one-dimensional original data by adding the first feature data and the restored one-dimensional second feature data G 3 , (S 46). Restores the two-dimensional or three-dimensional original image G by filling the first feature data G ⁇ "into a two-dimensional or three-dimensional space by pseudo-Hilbert scanning (S47).
• the restored original image G 0 is stored in the image memory 24 and output to the output means 25 such as a display (S48).
• FIG. 10 is a schematic diagram of an image information compression device and an image restoration device according to Embodiment 4 of the present invention.
• 1 is an image compression device
• 2 is an image decompression device
• 3 is an image input unit
• 4 is a frame memory
• 5 is a pseudo-Hilbert scanning unit
• 7 is a first encoding unit
• 8 is a second feature information extraction.
• Means 9 is the second encoding means
• 10 is the cumulative error memory
• 11 is the second feature data generation means
• 12 is the second feature image restoration means
• 13 is the judgment means
• 14 is the judgment table
• 15 is the base.
• Vector storage means 16 is an orthogonal transformation means, 17 is a quantization section, 18 is an entropy coding section, 19 is a first decoding section, 20 is a first characteristic image restoration section, 21 is a second decoding section, Reference numeral 22 denotes a second characteristic image restoring means, reference numeral 24 denotes an image memory, and reference numeral 25 denotes an output means. These are the same as those in FIG.
• the original image input from image input unit 3 is a moving image composed of a plurality of frame sequences that are continuous in the time direction.
• the first feature information extraction means 38 generates the first feature data by approximating the one-dimensional original data with a predetermined function having a greater redundancy than the one-dimensional original data for the predetermined frame. Further, the first feature information is extracted, and the first feature data is stored in the first feature data storage means 39. On the other hand, for the other frames, the first feature data of the past frame stored in the first feature data storage means 39 is subtracted from the one-dimensional original data to generate first difference data, and the first difference data is generated. The first feature information is extracted by generating the first feature data by approximating the first difference data with a predetermined function having greater redundancy than the data.
• the first encoding means 7 generates the first characteristic code data by subjecting the first characteristic data to event code, and the code amount determining means 41 determines the code amount of the first characteristic code data. Is calculated, and it is determined whether or not the code amount exceeds a predetermined threshold.
• the image restoration device 2 has an original image restoration means 42 and a first characteristic image storage means 43.
• the original image restoring means 42 generates a restored image by adding the first feature image generated by the first feature image restoring means 20 and the second feature image generated by the second feature image restoring means 22. If the restored image is the restored original image, the restored image is output to the image memory 24 as a restored original image, and the first feature image is stored in the first feature image storage means 43. . On the other hand, when the restored image is the restored original image, the original image restoring unit 42 adds the first feature image stored in the first feature image storage unit 43 and the restored image. The restored original image is output to the image memory 24.
• FIG. 11 is a flowchart showing an image information compression method according to Embodiment 4 of the present invention.
• the moving image is stored in the frame memory 4 (S51).
• the original image G x of a certain frame. Is stored in the frame memory.
• the pseudo-Hilbert running means 5 is the original image G x .
• the first feature information The extraction means 38 refers to the update flag held as an internal variable. here.
• the “update flag” is a flag for determining whether or not to update the first feature data stored in the first feature data storage unit 39. When this flag is in the ON state, The first feature data stored in the first feature data storage means 39 is updated.
• S is the original image G.
• the first feature information extracting means 38 stores the one-dimensional original data G xl in the first feature data storage means 39, and turns off the update flag (S55).
• the operation of the first feature information extracting means 38 in step S54 is the same as the operation in step S3 of the first embodiment, and thus a detailed description is omitted.
• the code amount judgment section 41 if the average step interval d is a predetermined threshold value gamma d less or scan Tetsupu density P is not smaller than a predetermined threshold value ⁇ ⁇ (S 56), the update flag to the ON state (S 57 ) o
• This is the first feature data when the temporal correlation of the image is lost due to a scene change of the moving image, or when the correlation between the current frame image and the past frame image is low. This is an operation for updating the first feature data stored in the storage means 39, and details will be described later.
• step S56 if the average step interval d is larger than the predetermined threshold value ⁇ d or the step density P is smaller than the predetermined threshold value ⁇ p ⁇ , the code amount determination unit 41 sets the update flag to the OFF state ( S 58).
• the first encoding means 7 encodes the first feature data G x1 , by entropy encoding. It is encoded (S59) and output as first feature code data (S60). At this time, a marker code indicating that the one-dimensional original data is encoded is added so that the first characteristic code data can be restored.
• the second feature information extracting means 8 and the second encoding means 9 encode the remaining second feature data G x2 , obtained by subtracting the first feature data G xl , from the one-dimensional original data G xl . , Is output as the second feature code data (S61 to S66), but since this operation is the same as steps S6 to S11 in the first embodiment, the description is omitted here.
• step S53 when the update flag is in the OFF state, the first feature extraction unit 38 acquires the past first feature data G yl , (y ⁇ x) stored in the first feature data storage unit 39. Then, the first feature data G yl , is subtracted from the one-dimensional original data G xl to generate first difference data G xl "( S68 ).
• step function changes stepwise, the first difference data G xl first feature data G xl approximating the "while converted into the cumulative square erroneous difference e a (p k, 1 k) the It is stored in the error memory 10 (S69).
• the first difference data G xl "minus the encoded child
• the first feature information having a small change in the time direction is subtracted, and the redundancy in the time direction is effectively removed, and as a result, it is possible to obtain high code efficiency for the first feature information.
• the information amount of the first feature information in the first difference data G xl "increases, so that the first feature data is increased to a certain extent in the first difference data G x , Past first feature data G yl , (y ⁇ x) stored in the data storage means 39 Need to be updated.
• the step in S 56 ⁇ S 58 by the code amount judgment section 41 monitors the average step interval d or step density P, the average step interval d is a predetermined threshold value gamma d less or step density p predetermined threshold gamma p When this is the case, it is determined that the correlation between frames in the time direction has decreased, and the update flag is set to the ON state, and the past first feature data G yl , ( y ⁇ x). This makes it possible to effectively remove redundancy in the time direction in encoding the first feature information.
• the original image G of the current frame There is judged whether the force not to correlate with the original image of the past frame is performed by the code amount determining unit 41 compares the average step interval d or step density p and the threshold value r d or gamma [rho, the update flag It was decided to switch (S56 to S58), but the original image G of the current frame.
• the determination of whether or not is correlated with the original image of the past frame is not limited to this method. For example, it is possible to make the determination based on whether or not the code amount Q of the first feature data G xl , exceeds a threshold value r Q.
• the code amount determination means 41 calculates the code amount Q of the first feature data G xl , and the code amount Q exceeds the threshold ⁇ ⁇ 3. It is also possible to determine whether the code amount Q exceeds a predetermined threshold value r Q and set the update flag to the ON state.
• step S57 the code amount determining means 41 first determines whether the current update flag is in the ON state or the OFF state.
• the update flag may be turned on. That is, as a result of generating the first feature data G xl , by approximating the first difference data G xl "in S69 , the average step interval d of the first feature data G xl , is equal to or less than a predetermined threshold r d or If the density p is not smaller than a predetermined threshold value gamma [rho, past frame of the original image stored with the original image G 0 of the current frame to the first feature data storage means 39 W 03
• the first feature data G may be generated by approximating the source data G.
• FIG. 12 is a flowchart illustrating a method of restoring an original image from hierarchical data encoded by the image information compression method according to the fourth embodiment.
• steps S71 to S77 are the same as those in steps S20 to S26 in FIG. 5 of the first embodiment, and a description thereof will not be repeated.
• step S77 the original image restoring means 42 sets the restored first feature image G x . , And the restored second feature image G x2 , are added to generate a restored image G x3 .
• the restored image G x3 has been subjected to the redundancy removal processing in the time direction, and therefore, it is necessary to extend the removed redundancy in restoration. Therefore, the original image restoring means 42 checks whether or not the first feature code data is added with a marker indicating that the marker is one-dimensional original data. If the marker is added (S78), the restoration is performed.
• the image G x3 is stored in the image memory 24 as a restored original image G x0 "(S79), and the restored first feature image G x0 , is stored in the first feature image storage means 43 (S 80).
• the restored original image G x0 "stored in the image memory 24 is output to output means 25 such as a display.
• the original image restoring means 42 returns the restored past past data from the first feature image storing means 43.
• the first feature image G yl , (y ⁇ x) is acquired (S82), and the first feature image G yl , and the restored image G x3 are added to generate a restored original image G x0 ", and the image memory 2
• the restored original image G x 0 "stored in the image memory 24 is Is output to output means 25.
• the image can be restored at high speed by a simple matrix calculation.
• the image information compression apparatus of each of the above embodiments may be configured to realize the image information compression apparatus of each of the embodiments by executing a program on a computer. Further, the program is recorded on a recording medium, and the program recorded on the recording medium is read by a computer and executed, thereby realizing the image information compression apparatus of each of the above embodiments. You may.
• the pseudo-Hilbert scan is used, so that the correlation of the pixel data along the scanning direction is large. Therefore, when the approximation is made by the first feature data, the contrast component and the edge component included in the image can be effectively separated from the texture component.
• the effective feature axis (base component) has a small image pattern dependence, and the test can be performed without calculating the base vector for each compressed image.
• the basis vector is calculated in advance using the pattern, and this is registered in the basis vector table, and can be used by reference.
• the coordinate values of the second feature image for each coordinate axis in the space spanned by the base vectors are significant because only the eigenvalues for some base vectors have significant values and the other eigenvalues are substantially zero.
• the second feature image can be compressed by approximating it with 0, leaving only the base vector with the significant value.
• the image information compression method according to the present invention can effectively separate an image for each feature, and can perform image information compression with high encoding efficiency and low image quality degradation. It is useful as a method for compressing image information in devices that transmit information or store image information, such as videophones and video cameras, and is particularly useful for image processing devices that require high speed and high coding efficiency. Are suitable.

Landscapes

• Engineering & Computer Science (AREA)
• Multimedia (AREA)
• Signal Processing (AREA)
• Compression Or Coding Systems Of Tv Signals (AREA)
• Compression, Expansion, Code Conversion, And Decoders (AREA)

Priority Applications (3)

Applications Claiming Priority (2)

Publications (1)

Family

ID=19188682

Family Applications (1)

Country Status (5)

Families Citing this family (15)

Citations (5)

Family Cites Families (11)

• 2001
  • 2001-12-25 JP JP2001392738A patent/JP3863775B2/ja not_active Expired - Fee Related
• 2002
  • 2002-01-09 KR KR20047008795A patent/KR100863018B1/ko not_active Expired - Fee Related
  • 2002-01-09 AU AU2002219564A patent/AU2002219564A1/en not_active Abandoned
  • 2002-01-09 US US10/498,818 patent/US7274826B2/en not_active Expired - Fee Related
  • 2002-01-09 WO PCT/JP2002/000038 patent/WO2003056836A1/ja not_active Ceased

Patent Citations (5)

Also Published As

Similar Documents

Legal Events