JP2013062644A - Image coding device and image decoding device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an image coding device having high coding efficiency and a corresponding image decoding device.SOLUTION: In an image coding device 100 including prediction means 11 and quantization means 15, and also inverse quantization means 18 and retention means 21, for coding with prediction applied to a unit block pixel, each quantization value in a signal channel constituting a pixel is divided into a reference signal channel for prediction on a quantization value and a predicted signal channel to which prediction is applied. In regard to a reference signal, the quantization value thereof is coded in advance and retained in quantization retention means 4. Based on the reference signal, quantization prediction means 5 and quantization compensation means 6 perform prediction and compensation on the quantization value of a predicted signal, to obtain a quantized prediction signal, so that the quantization prediction residual of a difference is coded. An image decoding device 300 performs decoding with corresponding processing.

Description

  The present invention relates to an image encoding device and an image decoding device, and more particularly to an image encoding device that predicts a quantization value to be encoded from a processed quantization value using inter-signal correlation and encodes a prediction error. And an image decoding apparatus.

  As methods for improving the coding efficiency in conventional image coding, there are a method for reducing temporal redundancy and a method for reducing spatial redundancy.

  As a method for reducing temporal redundancy, there are a frame difference method and a motion compensation method. The frame difference method simply subtracts two consecutive images and encodes the difference. In the motion compensation method, an approximate image of an encoding target frame is generated by applying a motion vector to a reference frame, and a difference from the encoding target frame is encoded. Since encoding is performed while reducing the difference between images, the motion compensation method is superior in encoding efficiency to the frame difference method. Various methods for estimating the motion vector used in the motion compensation method have been proposed, and the main prior art relating to motion estimation is introduced in Non-Patent Document 1.

  On the other hand, as a method for reducing spatial redundancy, there is a method of quantizing orthogonal transform coefficients. Orthogonal transformation maps pixel signals to the frequency domain and concentrates energy in the low range. By utilizing the fact that human visual characteristics are not sensitive to high frequencies, it is possible to increase coding efficiency by removing high frequency components by quantization. In addition, there is a method for reducing spatial redundancy by orthogonal transform coefficients or pixel prediction. Non-Patent Document 2 introduces main prior art related to orthogonal transform.

http://www.jpo.go.jp/shiryou/s_sonota/map/denki14/2/2-1-2-1.htm http://www.jpo.go.jp/shiryou/s_sonota/map/denki14/2/2-1-3.htm

Motion compensation reduces temporal redundancy but cannot be used for still images. On the other hand, the combination of orthogonal transformation and quantization and the spatial prediction method reduce the spatial redundancy, but none of them can reduce the redundancy of the color signal because the color signal is processed independently.

An object of the present invention is to solve the above-mentioned problems of the prior art and to provide an image encoding device with high encoding efficiency.

  Another object of the present invention is to provide an image decoding apparatus corresponding to the above-described image encoding apparatus with high encoding efficiency.

  In order to achieve the above object, the image encoding device of the present invention determines prediction information for predicting each pixel to be encoded from an encoded pixel for each pixel of a unit block including a plurality of pixels. A prediction unit that generates a prediction pixel of each pixel based on the prediction information, a difference unit that performs a difference process between the pixel and the prediction pixel to obtain a prediction residual, and the prediction A transforming unit that performs orthogonal transform on the residual to obtain a transform coefficient and a quantizer that quantizes the transform coefficient to obtain a quantized value and encodes the quantized value for each unit block In an image encoding device, quantization holding means for holding an encoded quantized value in a first signal channel constituting a pixel as a reference for prediction, and a code in a second signal channel constituting the pixel Quantum Quantization prediction means for determining quantization prediction information for predicting a value from a quantized value encoded in the first signal channel according to a predetermined prediction method, and a second signal channel based on the quantization prediction information A difference compensation process is performed between the quantization compensation means for generating a quantized prediction signal for the quantization value to be encoded and the quantization value to be encoded in the second signal channel and the quantized prediction signal. And a quantization difference means for making a quantized prediction residual, encoding information for specifying the predetermined prediction method, and encoding the quantized prediction residual for the second signal channel , The encoded quantized prediction residual and the quantized predicted signal are added to obtain the encoded quantized value in the second signal channel, and the encoded in the first and second signal channels. And generating the encoded pixels on the basis of Mino quantization value.

  In order to achieve the above object, an image decoding apparatus according to the present invention is an image decoding apparatus that generates decoded pixels by decoding code information encoded by the image encoding apparatus, and decodes the code information. Decoding means for obtaining a quantized value in the first channel, a quantized prediction residual in the second channel, and information identifying the predetermined prediction method, and a decoded quantum in the first channel A decoding-side quantization holding means for holding the quantized value as a reference for prediction, and a quantized value to be decoded in the second channel from the decoded quantized value in the first channel according to the predetermined prediction method Decoding side quantization prediction means for determining quantization prediction information for predicting the decoding, and decoding side for generating a quantized prediction signal in the second channel based on the determined quantization prediction information And a quantized prediction signal in the second channel by adding the quantized prediction signal in the second channel and the decoded quantized prediction residual to obtain a quantized value in the second channel. A decoded pixel is generated based on the decoded quantization value in the second signal channel.

  According to the image coding apparatus of the present invention, when the quantized value is encoded, the quantized value is encoded using the first signal channel as a reference for prediction, and the quantized value of the second signal channel is determined from the reference. Since the predicted quantized prediction signal is generated and the quantized prediction residual of the difference is encoded, the encoding efficiency can be improved.

  Further, according to the image decoding apparatus of the present invention, the code information encoded by the image encoding apparatus of the present invention can be decoded by performing a process corresponding to the process on the encoding apparatus side.

It is a functional block diagram of the image coding apparatus which concerns on embodiment of this invention. It is a functional block diagram of the image decoding apparatus which concerns on embodiment of this invention. It is a figure which shows typically the dependence relationship of the process sequence of a reference signal channel and a to-be-predicted signal channel. It is a detailed functional block diagram of the setting part by the side of an encoding apparatus. It is a figure for demonstrating the implicit setting of the reference | standard and to-be-predicted signal channel in a channel implicit setting part. It is a figure for demonstrating prediction and compensation of the quantization value in this invention.

  1 and 2 are functional block diagrams of an image encoding device and an image decoding device, respectively, according to an embodiment of the present invention. In FIG. 1, an image encoding device 100 includes a first switch 10, a prediction unit 11, a compensation unit 12, a difference unit 13, a conversion unit 14, a quantization unit 15, a channel-specific quantization value processing unit 16, an encoding unit. Means 17, inverse quantization means 18, inverse transform means 19, addition means 20, holding means 21, image input means 101 and code output means 102, and the pixel to be encoded is input from image input means 101, Each unit is processed by each unit or the like, and is output to the code output unit 102 as code information.

  Here, the quantization value processing unit 16 for each channel, which is a characteristic configuration of the present invention, includes the second switch 1, the quantization difference unit 2, the quantization addition unit 3, the quantization holding unit 4, the quantization prediction unit 5 and the like. , Quantization compensation means 6 and setting unit 7.

  In FIG. 2, an image decoding apparatus 300 includes a decoding unit 37, a (decoding side) channel-specific quantized value processing unit 36, a (decoding side) inverse quantization unit 38, a (decoding side) inverse transformation unit 39, and a decoding side. ) Addition means 40, (decoding side) holding means 41, (decoding side) compensation means 32, code input means 301 and image output means 302. Code information from the image encoding device 100 is input from the code input means 301. These units are processed for each unit block, and are output to the image output unit 302 as decoded pixels.

  Here, the channel-by-channel quantized value processing unit 36, which is a characteristic configuration of the present invention, includes a third switch 51, (decoding side) quantization adding means 53, (decoding side) quantization holding means 54, (decoding side). ) Quantization prediction means 55, (decoding side) quantization compensation means 56, and (decoding side) setting unit 57.

  Hereinafter, the processing of each means or each unit will be described with reference to the data flow shown in FIGS. Note that the image encoding apparatus 100 (appropriately abbreviated as an encoding apparatus) and the image decoding apparatus 300 (appropriately abbreviated as a decoding apparatus) have portions corresponding to processing. Such a portion will be described as appropriate by using () on the decoding device side, for example, as "compensation means 12 (compensation means 32)".

  When the encoded pixel to be referenced for prediction and compensation is held in the holding unit 21, the second switch 10 sends the pixel to be encoded from the image input unit 101 to the prediction unit 11 and the difference unit 13. Take the switch state to send. On the other hand, when the reference pixel is not held, such as the first unit block in the frame when the intra prediction is used, the state is such that the encoding target pixel is sent to the conversion unit 14.

  The prediction unit 11 refers to the encoded pixels held by the holding unit 21 and determines prediction information for approximating the image of the unit block to be encoded. The determined prediction information is sent to the compensation means 12 and the encoding means 17 (as indicated by the dotted line in FIG. 1).

  Various known techniques can be used to determine the prediction information. As an example, when using the intra-screen prediction of H.264, select the intra-screen prediction mode that individually encodes in each intra-screen prediction mode and minimizes the coding cost calculated from the code amount and the distortion amount, Predictive information. When using motion prediction, a frame and coordinates to be referenced are searched and used as prediction information.

  In the present invention, when calculating the coding cost in other functional blocks (particularly, each functional block of the setting unit 7 described later with reference to FIG. 4), it is similarly calculated from the code amount and the distortion amount. Shall.

  The compensation unit 12 (compensation unit 32) includes prediction information (prediction information decoded by the decoding unit 37) sent from the prediction unit 11 and encoded pixels (decoding) held by the holding unit 21 (holding unit 41). A predicted pixel of the unit block is generated. The generated prediction pixel is sent to the difference means 13 and the addition means 20 (addition means 40).

  The difference unit 13 determines the difference between the pixel to be encoded from the first switch 10 and the prediction pixel from the compensation unit 12 to obtain a prediction residual, and sends the prediction residual to the conversion unit 14.

  The conversion means 14 sends the prediction residual from the difference means 13 by sending the encoding target pixel from the first switch 10 to the prediction means 11 and the difference means 13 according to the switch state in the first switch 10. If it is, the prediction residual is converted, and if the pixel to be encoded is sent from the first switch 10, the pixel to be encoded is converted into a frequency domain transform coefficient by orthogonal transform. As the orthogonal transform, DCT (discrete cosine transform) to DCT approximate transform, DWT (discrete wavelet transform), or the like can be used.

  The quantization unit 15 quantizes the transform coefficient from the transform unit 14 to obtain a quantized value, and sends the quantized value to the second switch 1. Here, the quantization parameter used for quantization can be set as a combination of constant values. Further, the bit rate can be kept constant by controlling according to the information amount of the transform coefficient.

  In the channel-specific quantized value processing unit 16, the first switch 1 divides the quantized value from the quantizing unit 15 into each signal channel and then becomes a prediction reference provided by the setting unit 7 described later. According to a distinction between a channel (abbreviated as a reference signal channel) and a predicted signal channel (abbreviated as a predicted signal channel), the quantization value in the reference signal channel (as indicated by a one-dot chain line) and the quantization unit 17 The signal is sent to the holding means 4, and the quantized value in the signal to be predicted is sent to the quantization difference means 2.

  Here, each signal channel is a predetermined signal channel that constitutes a pixel to be encoded in the image input unit 101, and is a signal channel in each color space such as an RGB signal, a YUV signal, or a YcbCr signal. . In addition, any type and number of signal channels to be separated can be used. For the predetermined signal channel, for example, RGB, the setting unit 7 makes a distinction such that G is a reference signal channel and R and B are predicted signal channels.

  Note that the setting unit 7 sets the setting information regarding the distinction between the reference and the prediction target and other processing in the quantized value processing unit 16 for each channel. The setting unit 7 sends the setting information to the encoding unit 17 as indicated by a dotted line, and the encoding unit 17 encodes it. On the decoding device side, the setting information is decoded and set in the setting unit 57 as indicated by a dotted line, whereby the channel-specific quantization value processing unit 36 can perform processing corresponding to the processing on the coding device side. It becomes. Details of the setting information will be described later.

  The encoding unit 17 receives a quantization value in a reference signal channel described later, a quantized prediction residual in a predicted signal channel described later, setting information from the setting unit 7, and prediction information from the prediction unit 11. These are encoded and sent as code information to the code output means 102. For the encoding, a variable-length code or an arithmetic code that removes redundancy between codes can be used.

  The decoding unit 37 receives the code information encoded as described above from the code input unit 301, decodes it by the reverse process of the encoding unit 17, and determines the quantized value in the reference signal channel and the predicted signal channel. The quantization prediction residual, the setting information, and the prediction information are obtained, and the quantization value and the quantization prediction residual are supplied to the third switch 51, the setting information is set to the setting unit 57, and the prediction information is compensated by the compensation means 32. Send to.

  The third switch 51 sends the quantized value in the reference signal channel from the decoding unit 37 to the quantization holding unit 54 (as indicated by the alternate long and short dash line) by appropriately switching the switch state. The quantization prediction residual in the prediction signal channel is sent to the quantization addition means 53.

  The quantized value in the reference signal channel sent from the second switch 1 is encoded on the one hand by the encoding means 17 and, on the other hand, the quantized quantum signal for reference in the quantization holding means 4. Is stored as a digitized value. The encoded quantized value of the reference signal channel is held (as indicated by the alternate long and short dash line) and sent to the inverse quantization means 18 by the quantization holding means 4.

  The quantized value in the decoded reference signal channel sent from the third switch 51 is held as a decoded quantized value for reference in the quantization holding means 54. The decoded quantization value of the reference signal channel is held and input to the inverse quantization unit 38 by the quantization holding unit 54.

  The quantized value (quantized prediction residual) in the predicted signal channel sent from the second switch 1 (third switch 51) is the quantized value processing unit 16 for each channel (quantized value processing unit 36 for each channel). ), The following process is followed. The quantization difference means 2 takes the difference between the quantized value of the signal-under-predicted signal channel and the quantized prediction signal of the signal-under-predicted signal channel from the quantizing compensation means 6 to be described later. None, sent to the encoding means 17 and the quantization addition means 3 (processing only on the encoding device side). The quantized prediction residual is, on the one hand, sent to the encoding means 17 and encoded, and thus becomes an encoded quantized residual.

  The quantization addition means 3 (quantization addition means 53) is the quantization prediction residual from the quantization difference means 2 (third switch 51) and the quantization from the quantization compensation means 6 (quantization compensation means 56). By adding the predicted signal, the encoded quantized value in the predicted signal channel is obtained and sent to the quantization holding means 4 (quantization holding means 54).

  The quantization holding means 4 (quantization holding means 54) holds the encoded quantized value (decoded quantized value) in the predicted signal channel as well as the predicted signal channel, and reversely The data is sent to the quantization means 18 (inverse quantization means 38). Further, the quantization holding means 4 (quantization holding means 54) quantizes the encoded quantized values (decoded quantized values) held in both the predicted signal channel and the reference signal channel. The prediction means 5 and the quantization compensation means 6 (quantization prediction means 55 and quantization compensation means 56) are made available by reference.

  The quantization prediction means 5 quantizes the encoded reference signal channel held in the quantization holding means 4 for the quantization value to be encoded in the predicted signal channel sent from the second switch 1. Quantization prediction information to be predicted according to a predetermined prediction method is determined from the value and sent to the quantization compensation means 6. With the same function on the decoding device side, the quantization predicting unit 55 uses the quantized value of the encoded signal channel of the quantized prediction holding unit 54 to transmit the predicted signal channel transmitted from the third switch 51. Quantization prediction information for using the quantization prediction residual in the original quantization value is determined.

  The quantization compensation means 6 applies a predetermined prediction method based on the quantization prediction information from the quantization prediction means 5 and quantizes the encoding target in the predicted signal channel sent from the second switch 1. A quantized prediction signal, which is a value prediction signal, is generated and sent to the quantizing difference means 2 and the quantizing addition means 3. With the same function on the decoding device side, the quantization compensation unit 56 generates a quantized prediction signal based on the quantized prediction information from the quantization predicting unit 55 and sends it to the quantization adding unit 53.

  The inverse quantization means 18 (inverse quantization means 38) encodes the quantized values (decoded already) in both the reference signal channel and the predicted signal channel sent from the quantization holding means 4 (quantization holding means 54). The quantized value is converted into a transform coefficient by dequantizing the quantized value 15 in the reverse order of the quantization process in the quantizing unit 15. In addition, after the inverse quantization means 18 (inverse quantization means 38) after the channel-specific quantization value processing unit 16 (channel-specific quantization value processing unit 36), the processing returns to the common processing in each signal channel. The target data will be described by omitting the distinction between the reference and the predicted data.

  The inverse transform unit 19 (inverse transform unit 39) performs the inverse orthogonal transform by following the reverse procedure of the orthogonal transform in the transform unit 14, thereby obtaining the transform coefficient from the inverse quantization unit 18 (inverse quantization unit 38). No prediction residual is sent to the adding means 20 (adding means 40). In addition, when a pixel is sent from the first switch 10 to the conversion unit 14, it is not a prediction residual but a pixel.

  The adding means 20 (adding means 40) adds the prediction residual from the inverse transforming means 19 (inverse transforming means 39) and the predicted pixel from the compensating means 12 (compensating means 32), and performs the encoded pixel (decoding) Sent to the holding means 21 (holding means 41 and image output means 302). In the case of a pixel instead of a prediction residual as described above, the addition process is omitted and the pixel is sent to each means as it is.

  The holding unit 21 (holding unit 41) holds the encoded pixel (decoded pixel) from the adding unit 20 (adding unit 40) and uses it as a reference from the predicting unit 11 and the compensating unit 12 (compensating unit 32). Provide.

  The processing outline of each unit of the encoding device and the decoding device has been described above. FIG. 3 is a diagram schematically showing the dependence of the processing procedure on the reference and predicted signal channel in the processing. Here, a case will be described in which the signal channel is RGB, the reference signal channel is a G signal, and the predicted signal channel is a B signal and an R signal.

  Since the reference G signal is necessary to process the B signal and R signal to be predicted, it is necessary to encode in advance, and as shown in (1) by each process from the state of the pixel to be encoded It becomes a quantized value (as an output of the quantizing means 15), the quantized value is encoded as shown in (2), and can be referred to in the quantizing holding means 4 as an encoded quantized value. Become. It becomes possible to encode the B signal and the R signal to be predicted after the referenceable state.

  That is, the B signal to be predicted becomes a quantized value (as an output of the quantizing means 15) as indicated by (3) from each pixel to be encoded by each process, and the reference G signal as indicated by (4). By referring to the encoded quantization value in the signal, a quantized prediction residual is obtained as shown in (5), and the quantized prediction residual is encoded as shown in (6). For the predicted R signal, as shown in (7), (8), and (9), the quantized prediction residual is obtained by referring to the reference G signal as shown in (4), as shown in (4). Encoded.

  And, in the present invention, the processing as described in the above (1) to (9) is represented as (i-1) (i) (i + 1), so that the i-1th unit block, the ith unit block , The (i + 1) th unit block and each unit block are sequentially continued. During the continuation, the setting of the reference and predicted signal channel may be changed for each unit block. In the decoding apparatus, the predicted signal channel can be decoded only after the reference signal channel is decoded and the quantized value can be referred to.

  FIG. 4 is a detailed functional block diagram of the setting unit 7 on the encoding device side. The setting unit 7 includes an application setting unit 71, a channel explicit setting unit 72, a channel implied setting unit 73, and a channel learning setting unit 74. Each unit sets the processing method in the channel-specific quantization value processing unit 16 for each predetermined number of unit blocks, and the setting information is encoded by the encoding means 17. The encoding may be performed only at the beginning of a predetermined number of unit blocks, and may be reset at the switching location, that is, at the beginning when switching. . If the predetermined number of unit blocks are the entire image to be encoded by manual setting, the processing method can be set so as not to change the entire image. On the decoding device side, the setting information is decoded by the decoding unit 37 and the setting is read into the setting unit 57, whereby the channel-specific quantization value processing unit 36 performs the same process as the decoding device side.

  As described above, the application setting unit 71 sets whether or not to perform the process of distinguishing the reference and predicted signal channels. The setting may be a manual setting or may be automatically set based on the encoding cost. If not, processing in the reference signal channel is performed in all signal channels. That is, all quantized values of each signal channel are always sent as they are to the quantization holding means 4 and the encoding means 17 (quantization holding means 54) by the second switch 1 (third switch 51), and the quantization difference The processing of means 2, quantization addition means 3, quantization prediction means 5 and quantization compensation means 6 (quantization addition means 53, quantization prediction means 55 and quantization compensation means 56) is omitted, and the quantization prediction residual is Not calculated, only the quantized value is encoded in all signal channels.

  The channel explicit setting unit 72 explicitly sets the reference and predicted signal channels so that they need to be encoded, that is, for example, sets the R signal and the B signal as predicted using the G signal as a reference. Directly set and set. The setting may be manually given. Alternatively, a predetermined number of unit blocks may be individually encoded in each case of possible combinations of reference and predicted signal channels, and set to a combination that minimizes the encoding cost.

  For example, in the case of an RGB signal, refer to x for the remaining one z and the other two for reference x and y in three ways when the other two are to be predicted based on any one of RGB. The encoding costs may be compared by using 9 possible combinations of 6 ways (6 ways of assigning RGB to the xyz) as possible. The prediction y may be used as a reference for another predicted z so that y is predicted from the reference x and z is predicted from the y.

  The channel implied setting unit 73 sets the reference and predicted signal channels implicitly, that is, without being described directly in the setting information so that encoding is unnecessary. In this case, since a reference and a signal channel to be predicted are set by performing a given determination using encoded pixels, information that specifies how to make the given determination instead of directly describing the channel. Is included in the setting information, and the same setting as that on the encoding device side can be used by making the same determination on the decoding device side.

  As the given judgment, the richness (degree of change) of the quantized value (not the pixel) in the encoded block in the neighborhood region of the encoding target unit block, as shown in FIG. There are criteria and prediction decisions based on comparisons in the signal channel. In other words, in the example of the RGB signal shown in FIG. 5, a predetermined code in block units having a common shape and a common relative position with respect to the encoding target blocks Ri, Gi, Bi at the same time and the same position in each signal frame. The reference and the prediction target are determined by comparing the abundance of changes in the quantized values of the normalized neighborhood regions Rs, Gs, and Bs.

  As a criterion for whether or not there is a large amount of change, the absolute value variance or absolute value amplitude of the coded quantized value in each neighboring region can be used, and the larger the absolute value variance or amplitude, the richer the change. . If the comparison is made so that signal channels with abundant changes are adopted as a reference, the effect of improving the prediction accuracy can be obtained. As a result of comparison, if it is set so that a signal channel rich in change is adopted as the prediction target, the effect of improving the coding efficiency can be obtained. As in the example of the channel clarification unit 72, a predetermined combination is set as to which of the signal channels having the highest level of change is used as the reference or prediction from which of the lower signal channels. .

  The channel learning setting unit 74 receives the quantized value of each channel signal of the encoding target block as an input, and automatically outputs the setting information of the reference and predicted signal channel in the encoding target block. It becomes. Here, the input quantization value is prepared by the same processing as that for the reference signal. The automatic output is performed by using a classifier (for example, a decision tree, a random forest, a neural network, or the like) based on predetermined statistical machine learning.

  That is, with a given quantized value as a feature quantity in a given learning image and a given setting of a reference and a predicted signal channel that minimizes the coding cost for that learning image, A classifier based on statistical machine learning is constructed, and setting information is automatically output using the classifier. The channel learning setting unit 74 further obtains a setting that minimizes the coding cost for each predetermined unit block (regardless of the discriminator), and compares the discriminator with the output of the discriminator. Sequential learning may be applied. Note that since the decoding device only receives explicit reference and prediction setting information, the information such as the statistical machine learning and the discriminator used need not be encoded.

  Next, a predetermined prediction method for predicting and compensating the quantized value of the predicted signal channel from the quantized value of the reference signal channel in the quantizing predicting means 5 and 55 and the quantizing compensating means 6 and 56 This will be described with reference to FIG. The prediction method is also set by the setting unit 7, encoded as setting information, and the same processing is performed on the decoding device side.

  In FIG. 6, as an illustrative example, the G signal in the RGB signal is used as a reference, and the B signal is used as a prediction target. As shown in (1) and (2), the quantized values of the reference G signal and the predicted B signal are both determined by the details of the conversion method to the conversion coefficient by the conversion means 14 and are a predetermined size (here, 4 × 4). The frequency elements are sequentially encoded according to a predetermined order, such as a zigzag scan starting from a direct current element (DC). In addition, in order to distinguish from the unit block which consists of a pixel instead of a frequency domain, it will call a "matrix" and an "element".

  As indicated by a thick frame F1 in (1), all elements in the matrix have already been encoded for the reference G signal. Here, the elements of the portion surrounded by the thick frame F2 in (2) are sequentially predicted and compensated according to the predetermined order and encoded, but the remaining portion is the predicted B signal before encoding. A prediction method will be described using the next prediction / compensation and encoding target element B (x). Note that in the notation B (x), G ′ (x), etc., x represents a position in the matrix, and represents whether it has been encoded or not, with or without a comma.

  B (x) is each encoded element in the reference signal surrounded by F1, and each encoded element at the time of encoding B (x) in the predicted signal surrounded by F2. Prediction / compensation is performed by applying a predetermined prediction formula to the above. The parameter for applying the prediction formula is the quantized prediction information determined by the quantization prediction means 5 and 55, and the quantization compensation means 6 and 56 apply the parameter to the prediction formula to obtain the quantized prediction signal. obtain.

  For example, in particular, a prediction formula of B (x) = k × G ′ (x) may be used by multiplying an encoded element G ′ (x) at the same position in the reference signal by a coefficient k as a parameter. Furthermore, for example, the following equation (Equation 1) or (Equation 2) can be used as a prediction equation for B (x) as a prediction equation in which the coefficient k is embodied.

Here, G ′ (x _i ) and B ′ (x _i ) are respectively the quantized value of the reference G signal at position x _i and the predicted B signal as shown in the figure where i = 1 and 2 Represents the encoded quantized value of. The position x _i represents each position in a predetermined neighborhood area that has been encoded with respect to the position x of the encoding target element in the predicted B signal. n represents the number of elements belonging to the neighboring region. median ⁿ _i represents the median value of n. w _i is a weighting coefficient, and can be set according to the distance between the positions x _i and x. The predetermined neighboring region in the matrix is an example which is constructed according to the next x ₁ and left x ₂ on the encoding target position x 6 may be set to any other shape .

That is, (Equation 1) and (Equation 2) are both ratios B (x _i ) / at predetermined positions x _i in a predetermined neighborhood region in the matrix as a quantity expressing the inter-signal correlation. Focusing on G (x _i ), the coefficient k is determined.

However, when the element x is the first position (DC position) in the matrix or when all G ′ (x _i ) are zero, the quantized prediction signal B (x) in the element is set in advance. It is a fixed value. Alternatively, when some G ′ (x _i ) is zero, it is excluded from the calculation formula.

As a calculation example, it is assumed that a predetermined neighborhood region with respect to the encoding target position x is composed of x ₁ and x _{2 in} FIG. 6, and B (x ₁ ) = 10, B (x ₂ ) = 25, G (x ₁ ) = 20, G (x ₂ ) = 30, w ₁ = w ₂ = 1/2, the coefficient k is (Equation 1)
k = median (10/20, 25/30) = 1/2 or 5/6 [The median value is set to either because it is an even number]
When (Formula 2) is adopted,
k = (10/20 + 25/30) / 2 = 2/3
It becomes. The parameter k multiplied by the value of G ′ (x) at the same position as B (x) is the quantized prediction signal for B (x).

  In addition, various modified embodiments will be described. The setting of processing in each modified embodiment (including processing other than the channel-specific quantization value processing units 16 and 36) is also encoded as setting information by the setting unit 7, and the same processing can be performed on the decoding side.

  The quantization prediction means 5 and 55 may determine the quantization prediction information for only a predetermined part of the matrix in which each frequency element is listed. In this case, the quantized prediction residual is obtained and encoded only for the predetermined part. The remaining portion is a signal channel to be predicted, but the quantized prediction residual is not obtained, and the quantized value may be encoded or deleted as in the reference signal channel. As the predetermined part, a part of a predetermined frequency region in the matrix or a part larger than a certain reference value can be adopted. The setting information is set by the setting unit 7, then encoded, and the same is performed on the decoding device side.

  In the quantization difference means 2, the quantization prediction residual may be set to an integer value by performing a rounding process as an additional process after the difference process. In this case, since the quantized prediction residual is also an integer like the normal quantized value, the effect of improving the coding efficiency can be obtained.

  The quantization means 15 may omit the rounding process in the quantization and perform only the division by the quantization step. In this case, each means treats the quantized value described above as a real number instead of an integer, but the processing is common. In addition, in the dequantization means 18 and 38, only integration by a quantization step is performed. In this case, in particular, by simultaneously applying the rounding process in the quantization difference means 2 described above, both the quantized value and the quantized prediction residual become real numbers, and the effect of improving the prediction accuracy is obtained.

  As described above, according to the present invention, an input signal is separated into quantized values in a plurality of signal channels, and other quantized values are predicted from the quantized values serving as reference signals to reduce the amount of information generated in the predicted signal. By including the means to perform, high encoding efficiency is possible. Further, it can be combined with a conventional prediction method that reduces temporal redundancy and the like, and the encoding efficiency can be further improved.

  DESCRIPTION OF SYMBOLS 100 ... Image coding apparatus, 300 ... Image decoding apparatus, 10 ... First switch, 11 ... Prediction means, 12, 32 ... Compensation means, 13 ... Difference means, 14 ... Conversion means, 15 ... Quantization means, 16, 36: Quantized value processing unit for each channel, 17: Coding means, 37: Decoding means, 18, 38 ... Inverse quantization means, 19, 39 ... Inverse transform means, 20, 40 ... Addition means, 21, 41 ... Holding Means 101: Image input means 102: Code output means 301 ... Code input means 302 ... Image output means 1 ... Second switch 51 ... Third switch 2 ... Quantization difference means 3, 53 ... Quantization addition means, 4,54 ... Quantization holding means, 5,55 ... Quantization prediction means, 5,56 ... Quantization compensation means, 7,57 ... Setting section

Claims (13)

Prediction means for determining prediction information for predicting each pixel to be encoded from encoded pixels for each pixel of a unit block composed of a plurality of pixels;
Compensation means for generating a prediction pixel of each pixel based on the prediction information;
Difference means for performing a difference process between each pixel and the prediction pixel to obtain a prediction residual,
Transform means for transforming the prediction residual into a transform coefficient by performing orthogonal transform;
In an image coding apparatus that includes a quantization unit that quantizes the transform coefficient to obtain a quantized value, and performs coding of the quantized value for each unit block.
Quantization holding means for holding an encoded quantized value in a first signal channel constituting a pixel as a reference for prediction;
Quantization prediction for determining quantization prediction information for predicting a quantization value to be encoded in a second signal channel constituting a pixel from a coded quantization value in the first signal channel according to a predetermined prediction method Means,
Quantization compensation means for generating a quantized prediction signal for the quantization value to be encoded in the second signal channel based on the quantization prediction information;
A quantization difference means for performing a difference process between the quantization value to be encoded in the second signal channel and the quantized prediction signal to obtain a quantized prediction residual;
The information specifying the predetermined prediction method is encoded, and the quantization prediction residual is encoded for the second signal channel, and the encoded quantization prediction residual and its quantization prediction are encoded. The signal is added to form an encoded quantized value in the second signal channel, and the encoded pixel is generated based on the encoded quantized value in the first and second signal channels. An image encoding apparatus characterized by that. Further comprising an application setting unit for setting the presence / absence of application of the quantization prediction means, quantization compensation means, and quantization difference means for each predetermined number of unit blocks;
2. The image code according to claim 1, wherein the information about the presence / absence of application is encoded, and the quantized value is encoded in both the first and second signal channels when there is no application. Device.   It further comprises a channel explicit setting unit that sets the first and second signal channels for each predetermined number of unit blocks, and encodes information of the set first and second signal channels. The image encoding device according to claim 1.   The channel explicit setting unit may determine the first and second combinations based on the encoding cost in each case of the possible combinations of the first and second signal channels so that the encoding cost is minimized. The image coding apparatus according to claim 3, wherein a signal channel is set.   A channel implied setting unit configured to set the first and second signal channels for each predetermined number of unit blocks, the channel implied setting unit configured to change each of the degree of change in the quantized value in a predetermined coded neighborhood region; The information of the first and second signal channels to be set is not encoded by setting the first and second signal channels based on a comparison between the signal channels. The image encoding device described. A channel learning setting unit that sets the first and second signal channels for each predetermined number of unit blocks, and encodes information of the set first and second signal channels;
The channel learning setting unit is configured for given learning data including a given learning image and a given setting of first and second signal channels that minimize the coding cost for the learning image. 2. The image coding apparatus according to claim 1, wherein the first and second signal channels are set by using a discriminator obtained by applying a given statistical machine learning. The quantized values are configured in a matrix of a predetermined size in the frequency domain for each unit block, and each element of the matrix is encoded according to a predetermined order,
The quantization predicting means, according to the predetermined prediction scheme, sequentially converts each element to be encoded in the matrix in the second signal channel in the matrix in the first and second signal channels. The image coding apparatus according to claim 1, wherein the quantization prediction information is determined as information for prediction based on a predetermined encoded element. The quantization predicting means, according to the predetermined prediction method, sequentially encodes each element to be encoded in the matrix in the second signal channel in the matrix in the encoded first signal channel. Predict by multiplying the element at the same position by a constant,
The constant is the value of the element of the second signal channel at each position in the predetermined neighborhood region encoded in the matrix of the first and second signal channels. The ratio divided by the element value is calculated as a median value in the neighboring region or a predetermined weighted average in the neighboring region, and the constant is determined as the quantized prediction information. Item 8. The image encoding device according to Item 7.   The quantization prediction means determines the quantization prediction information for only a predetermined part of the matrix, so that the quantization prediction residual for only the predetermined part for the second signal channel is determined. 9. The encoding method according to claim 7, further comprising: encoding the quantized value for a portion other than the predetermined portion, and encoding information for specifying the predetermined portion. Image coding apparatus.   10. The quantization prediction unit sets, as the predetermined part, a predetermined low frequency region in the matrix or a region in which the absolute value of an element in the matrix satisfies a predetermined criterion. The image encoding device described in 1.   The image code according to any one of claims 1 to 10, wherein the quantization difference means obtains the quantization prediction residual as an integer value by performing a rounding process after performing the difference process. Device.   The image coding apparatus according to claim 1, wherein the quantization unit omits a rounding process in quantization so that each unit handles the quantized value as a real value. An image decoding device that generates decoded pixels by decoding code information encoded by the image encoding device according to any one of claims 1 to 12,
Decoding means for decoding the code information to obtain a quantized value in the first channel, a quantized prediction residual in the second channel, and information specifying the predetermined prediction method;
Decoding side quantization holding means for holding the decoded quantization value in the first channel as a reference for prediction;
Decoding-side quantization prediction means for determining quantization prediction information for predicting a quantization value to be decoded in the second channel according to the predetermined prediction method from the decoded quantization value in the first channel;
Decoding side quantization compensation means for generating a quantized prediction signal in the second channel based on the determined quantized prediction information,
The quantized prediction signal in the second channel and the decoded quantized prediction residual are added to obtain a quantized value in the second channel, which is decoded in the first and second signal channels. An image decoding apparatus that generates a decoded pixel based on a quantized value.