JP3024785B2 - Orthogonal transformation method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to an orthogonal transform method used for encoding and decoding audio, still images, moving images, and the like.

(Prior Art) Conventionally, as a technique in this kind of field, for example, there has been the following.

Reference 1: “A lapped-orthogonal-transfom based
variable bit−rate video coder for packet
networks "IEEE ICASSP'89 (1989) (US) P. 1905-1908 Reference 2:" Transfom / subband coding of speech wi
th the lapped orthogonal transfom "IEEE ISCAS '
89 (1989) (U.S.A.) P. 1268-1271, Lapped Orthogonal transfom (hereinafter, referred to as
There is a method of transforming an image signal into a frequency domain by LOT transformation, and quantizing and encoding the transform coefficient. The LOT transform is one of orthogonal transform methods. Compared to the conventionally used Discrete Cosine Transform (hereinafter, referred to as DCT transform), quantization distortion of coefficients is less noticeable, and even if there is a coding error. It has the advantage that image deterioration is less noticeable, and has recently attracted attention.

Reference 2 discloses a method for creating a transformation matrix for LOT transformation. According to Document 2, the LOT transform of a one-dimensional signal and a two-dimensional signal is described as follows.

1 dimensional signal, y = P ^t x (conversion) ^{x ~} = Py (inverse transform), and the two-dimensional signal, Y = P ^t XP (conversion) ^{X ~} = PYP ^t (inverse transform) However, x; 2M Dimensional input signal vector y; M-dimensional conversion coefficient vector x ^to ; 2M-dimensional reproduction signal vector X; 2M × 2M input signal matrix Y; M × M conversion coefficient matrix X ^to ; 2M × 2M reproduction signal matrix P; 2M × M LOT Transformation matrix P ^t ; transposed matrix of M × 2M LOT transformation matrix.

Here, the reason why even if there is a code error in the LOT transform, the code error is not conspicuous will be described. Incidentally, an input signal x, and the transform coefficients y, the output signal x ^~ as will be described one-dimensional LOT transform in order to simplify the description.

The LOT transformation and the inverse transformation of a certain one-dimensional signal (for example, an audio signal) are shown in FIGS.

In FIG. 2, an input audio signal X (FIG. 2A)
From the input signals x _{1 to} x ₅
Is obtained ((b) in the same figure). Subsequently, it was LOT transform seeking transform coefficients y ₁ ~y ₅ (FIG. (C)), the transform coefficients y ₁
When subjected to LOT inverse transformation ~y _5, the inverse transform output x ₁ ^~ ~x ₅ ^~ a local reproduction signal is obtained (FIG. (D)). The LOT transform is an orthogonal transform in which a final reproduced signal is obtained only by adding the local reproduced signals ((e) in the figure).

Here, the conversion coefficient y _3, if missing by the error code, the local reproduction signals x ₂ ^- and x ₄ Effect of addition of ^~
The final output as shown in FIG. However, if a similar coding error occurs in M × M DCT transform coding, the signal is completely lost as shown in FIG. As described above, the LOT transform has a feature that a large quality deterioration does not occur at the time of a coding error and the LOT transform is resistant to the coding error.

The above document 1 describes an image packet encoding method using LOT transform, taking advantage of the feature that LOT transform is resistant to encoding errors. In the packet communication method, encoded audio data and image data are transferred into packets (cells).
And a header indicating the destination and the type of the content is added, and large-capacity communication is performed by transmitting and receiving the cell. In this packet communication method, if a cell exceeding the processing capacity of the network is sent, an overflowing cell is discarded. In this document 1, LOT transform is used for transform coding, and coding error and cell discarding are performed. We propose an image coding method that is strong against

(Problems to be Solved by the Invention) However, in the orthogonal transform method having the above-described configuration, when an input signal having a finite length is transformed and coded by the LOT transform to completely reproduce the input signal, the DCT uses the input sample for the input sample. However, the LOT transform has a problem that the number of transform coefficients is slightly increased with respect to input samples, and the coding rate is increased. As shown in FIG. 3, for example, for a one-dimensional signal X (64 samples)
When subjected to 16 following LOT conversion, 16-dimensional input signal vector cut can nine (x ₁ ~x _9). 8 for each input vector
Since the dimensional transformation coefficient vector (y _{1 to} y ₉ ) is obtained, the number of transformation coefficients is 72 (= 8 samples × 9), which is larger than 64 input samples.

Also, when the same number of transform coefficients as the number of input samples or less are sent, there is a problem that the beginning and end of the output signal deteriorate. That is, if the transform coefficients y ₁ and y ₉ shown in FIG. 3 are not coded, the corresponding reproduced signal portion deteriorates. This is the same in the case of the two-dimensional LOT transform.

An object of the present invention is to provide an orthogonal transform method that solves the problems of the prior art that the coding rate increases and that the reproduction signal deteriorates.

(Means for Solving the Problems) According to a first aspect of the present invention, a plurality of signal blocks are generated by sequentially extracting a finite-length input signal, and a LOT is performed on each of the signal blocks. In the orthogonal transformation method for performing transformation and encoding the input signal, the input signal is set in advance so that a start point and an end point of the input signal are continuous, and the cutout is performed.

In the second invention, a plurality of signal blocks are generated by sequentially cutting out a finite-length input signal, and a plurality of transform coefficients are obtained by performing LOT transform on each of the signal blocks. LOT inverse transform is performed to generate a plurality of local reproduction signals, and in the orthogonal transformation method of reproducing the input signal by adding the respective local reproduction signals, among the respective local reproduction signals, of the first local reproduction signal, The reproduction is performed by adding the start point side to the end point side of the last local reproduction signal.

(Operation) Since the first invention constitutes the orthogonal transformation method as described above, for example, an audio signal, an image signal, or a prediction residual signal of an audio signal or an image signal is converted by an LOT transform into a finite-length input signal. In encoding, the starting point and the ending point of the input signal are treated as being continuous, signal blocks for encoding are sequentially cut out from the input signal, and each signal block is subjected to LOT transform to obtain a plurality of transform coefficients. . As a result, the number of signal blocks to be cut can be reduced by one, and the same number of transform coefficients as the input signal can be obtained.

According to a second aspect, a plurality of local reproduction signals are generated by performing an LOT inverse transform on each transform coefficient obtained in the first aspect. When reproducing the input signal by adding the respective local reproduction signals,
The reproduction is performed by adding the start point of the first local reproduction signal to the end point of the last local reproduction signal. Therefore, the above problem can be solved.

(Embodiment) FIG. 1 is a block diagram showing the configuration of an image processing apparatus for implementing an orthogonal transform method according to a first embodiment of the present invention.

This image processing apparatus has a television camera 51 for converting an input original image Si into an analog electric signal.
Is connected to an analog / digital converter (hereinafter, referred to as an A / D converter) 52. The A / D converter 52 has a function of converting an analog electric signal into, for example, an 8-bit digital image signal Dg.

Further, the output side of the A / D converter 52 is connected to a QMF 53 having an analysis filter function for band division.
Are connected to encoders 54a, 54b, 54c, 54d for intra-frame prediction coding and motion compensation inter-frame coding + LOT, respectively. And these encoders 54a, 54b, 54c,
The output side of 54d is connected to the multiplexer 55, and the multiplexer 55
Are connected via a transmission line 56 to a separator 57 for data separation.

On the output side of the separator 57, decoders 58a, 58b, 58c, 58d such as intra-frame prediction decoding and motion compensation inter-frame decoding + LOT are provided.
Are connected, and a QMF 59 having a function as a synthesis filter is connected. The output side of QMF59 is a digital / analog conversion digital / analog converter.
An analog converter (hereinafter, referred to as a D / A converter) 60 and an image monitor 61 are sequentially connected.

FIG. 4 is a block diagram showing the internal configuration of the encoder 54a in FIG. Note that encoders 54b to 54d have the same configuration as encoder 54a.

The encoder 54a is configured by an intra-frame encoder, has an input terminal 54a-1 for inputting an output of the QMF 53, and has an input terminal 54a-1 and an output terminal 54a-9.
Buffer 54a-2 and LO for buffer 54a-2 output
LOT converter 54a-3 that performs T conversion and outputs a conversion coefficient
And a quantizer 54a- for quantizing the output of the LOT transform 54a-3.
4 are sequentially connected.

Further, on the output side of the quantizer 54a-4, an inverse quantizer 54a-5 for inversely quantizing the transform coefficient and an inverse quantizer 54a
LOT inverse transform converter 54a-6 which performs an LOT inverse transform on the output of −5
, A buffer 54a-7 and a frame memory 54a-8 for storing a local reproduction signal are sequentially connected. here,
The buffers 54a-2 and 54a-7 have a function of fetching image data of one frame and performing a predetermined data rearrangement on the image data.

FIG. 5 is a block diagram showing the internal configuration of the decoder 58a in FIG. The decoders 58b to 58d have the same configuration as the decoder 58a.

The decoder 58a is composed of an intra-frame decoder, and has an input terminal 58a-1 for code input and an output terminal 58a-2.
Between the input and output terminals 58a-1 and 58a-2, an inverse quantizer 58a-3, an LOT inverse transformer 58a-4, and a buffer 58a.
-5 is cascaded.

The orthogonal transformation method of the image processing apparatus configured as described above will be described separately for (A) encoding and (B) decoding.

(A) At the time of encoding, the input original image is obtained by scanning the imaging tube of the TV camera 51.
The signal is converted into an electric signal indicating the light and shade distribution, and the A / D converter 52
, The pixel is 8-bit quantized and sampled,
It is converted to a digital image signal. The digital image signal is divided into four bands by the QMF 53, and the image signals in each band are encoded by the encoders 54a to 54d as follows.

First, when an image signal is input to the input terminal 54a-1 shown in FIG. 4, the image signal is taken into the buffer 54a-2 for one frame, the data is rearranged, and then the LO
LOT conversion is performed by the T converter 54a-3. The LOT coefficient is quantized and dequantized by the quantizer 54a-4 and the dequantizer 54a-5, respectively, and encoded. Inverse quantizer 54a-5
The LOT inverse transform is performed by the LOT inverse transformer 54a-6, and the output transform coefficient is stored in the buffer 54a-7 for one frame.
Further, the output image is rearranged and added in the buffer 54a-7, and is added to the frame memory 54a-8.
Is stored as a local reproduction signal. On the other hand, the quantizer 54a
The LOT coefficient encoded by -4 is output from the output terminal 54a-9.

As described above, the LOT coefficients encoded by the encoders 54a to 54d are multiplexed by the multiplexer 55 and transmitted to the transmission path 56.

(B) Decoding At the time of decoding, the code sent from the transmission path 56 is separated into codes of each band by the separator 57 and input to the decoders 58a to 58d. These decoders 58a to 58d reproduce the image signals of the respective bands as follows.

The code input from the input terminal 58a-1 is the inverse quantizer 5
Inverse quantization is performed in 8a-3. LOT coefficient dequantized is LOT
After the inverse conversion is performed by the inverse converter 58a-4, the data is stored in the buffer 58a-5 for one frame, and the data is rearranged.
The added signal is output from the output terminal 58a-2 as a reproduced signal.

As described above, the decoded outputs of the decoders 58a to 58d are input to the QMF 59. The QMF 59 combines the four band-divided image signals as one reproduced image signal. This QMF59
Is returned to an analog image signal by the D / A converter 60 and output to the image monitor 61.

Next, the rearrangement or addition of data performed in the buffers 54a-2, 54a-7 and 54a-8 will be described with reference to FIG.

FIG. 6 is an explanatory diagram for explaining the orthogonal transform method of the present invention. For simplicity, the input signal is assumed to be a one-dimensional signal X (64 samples).

As shown in FIG. 6, the input signal is rearranged by arranging the same signal X as the one-dimensional signal X before and after the one-dimensional signal X, and presetting so that the start point and the end point of the signal are continuous. deep. Here, considering a one-dimensional LOT using a 16 × 16 transformation matrix, if the one-dimensional signal X is cut out by overlapping eight samples by 16 samples, the data to be cut out is x _{1 in} all.
You can do _nine up to x9. However, since the x ₁ ~x ₉ it is the same, the transform coefficients will be be a eight sets of transform coefficients y ₁ ~y _8, the 64 transform coefficients in total.

When playing reproduces from the local reproduction signals x ₁ ^- x ₈ to ^~ the transform coefficients y ₁ ~y ₈ and LOT inverse transform, sorted from the local reproduction signals x ₁ ^- x ₈ to ^~, local the reproduction signal x ₁ 8 samples before ^half-adding half 8 sample after ^~ x _8.

By doing so, the transform coefficients can be saved by one block, and the input signal X can be completely reproduced with the same number of transform coefficients as the input samples.

In this embodiment, since subband coding using QMF is performed, coding is performed by allocating a large number of coded bits for a low-frequency signal that is visually important, and coding is performed with a small number of coded bits for a high-frequency signal. Make As a result, a reduction in the coding rate can be expected.

7 (a) and 7 (b) are block diagrams showing the main parts of an image processing apparatus according to a second embodiment of the present invention. FIG. 7 (a) shows the configuration of the encoder 54a shown in FIG. FIG. 2B is a block diagram showing the internal configuration of the decoder 58a shown in FIG.

7 (a) is used in place of FIG. 4, and FIG. 7 (b) is used in place of FIG. Note that the encoders 54b-5
4d and the decoders 58b to 58d are also the encoder 54a and the decoder 5
Each has the same configuration as 8a.

In FIG. 7 (a), this encoder is composed of a motion-compensated inter-frame prediction encoder and has an input terminal 54a-A for inputting an output of the QMF 53, and the input terminal 54a-A.
A includes adders 54a-B, buffers 54a-C, LOT converter 5
4a-D and quantizers 54a-E are sequentially connected.
Further, the output side of the quantizer 54a-E is connected to the output terminal 54a-F, and the output side of the quantizer 54a-E and the input terminal 54a-F.
-A, an inverse quantizer 54a-G, a LOT inverse transformer 54a
-H, a buffer 54a-I, an adder 54a-J, a frame memory 54a-K, and a motion vector detector 54a-L are sequentially connected. The motion vector detector 54a-L has a function of matching the image signal one frame before the frame memory 54a-K and detecting the motion vector of the output image of the QMF 53.

Further, the output side of the motion vector detector 54a-L is commonly connected to the frame memories 54a-K and the output terminals 54a-F,
The input side of the motion vector detector 54a-L and the adders 54a-B, 54
The input side of a-J is commonly connected.

In FIG. 7 (b), this decoder is composed of a motion-compensated inter-frame predictive decoder and has an input terminal 58a-A for inputting a code and an output terminal 58a-B for outputting a reproduced signal. An inverse quantizer 58a-C, an LOT inverse transformer 58a-D, a buffer 58a-E, and an adder 58a-F are sequentially connected between the input and output terminals 58a-A and 58a-B. Further, the output terminal 58a-B and the input terminal 58a-A are connected to the frame memory 58a.
-G is connected to the input side thereof and its frame memory 58a-G
Are connected to adders 58a-F.

 Next, the operation will be described.

When the output of the QMF 53 in FIG. 1 is input to the input terminal 54a-A, the motion vector detector 54a-L matches the signal one frame before in the frame memory 54a-K to detect the motion vector. Is done. Subsequently, the adder 54a
According to -B, the prediction signal considering the motion vector is subtracted to be a residual signal. This residual signal is
After one frame is stored in 54a-C and the signals are rearranged, LOT conversion is performed by the LOT converters 54a-D, and conversion coefficients are output. The output transform coefficient is output to the quantizer 54a.
-E and quantization by the inverse quantizers 54a-G, respectively.
It is inversely quantized and encoded.

Further, the output transform coefficients of the inverse quantizers 54a-G are subjected to the LOT inverse transform by the LOT inverse transformers 54a-H, and the buffer 54a-G
One frame is stored in I. After the data is rearranged in the buffer 54a-I, addition is performed. The predicted signal subtracted earlier is added by the adder 54a-J, and is stored in the frame memory 54a-K as a local reproduction signal. The sign of the coefficient and the motion vector are output to an output terminal 54.
aF to the multiplexer 55 shown in FIG.

The symbols from the input terminals 58a-A shown in FIG.
The motion vectors are separated and inversely quantized by inverse quantizers 58a-C. The output transform coefficients of the inverse quantizers 58a-C are L
The LOT is inversely transformed by the OT inverse converter 58a-D,
One frame is stored in a-E, the data is rearranged, and added. After that, a prediction signal considering the motion vector is added by the adders 58a-F, and the frame memory 58
a-G is stored as a reproduction signal. The obtained reproduction signal is output from output terminals 58a-B.

Note that the present invention is not limited to the illustrated embodiment, and various modifications are possible. For example, there are the following modifications.

(I) In the above embodiment, the subband coding by QMF has been described, but the QMFs 53 and 59, the multiplexer 55 and the demultiplexer 57 are used.
, And furthermore, a plurality of configured encoders 54a to 54d
And one of the decoders 58a to 58d, respectively, and the same effect can be expected in this case as well.

(II) When performing two-dimensional LOT transformation on an image signal, use the eighth
As shown in the figure, if the input image signals are spread, the same effect as the one-dimensional LOT transform can be obtained. Note that Q1 to Q3 in FIG. 8 are signal blocks for extraction.

(Effects of the Invention) As described above in detail, according to the first invention, the input signal is set in advance so that the start point and the end point of the input signal are continuous, and then the input signal is cut out to perform the LOT conversion. As a result, one signal block to be cut can be saved, and the same number of transform coefficients as the input signal can be obtained. As a result, the coding rate can be reduced, and the deterioration of the reproduced signal can be prevented.

In the second invention, when the transform coefficient obtained in the first invention is inversely transformed to obtain a reproduced signal, after the LOT inverse transform, the start point side of the first local reproduced signal is changed to the end point of the last local reproduced signal. Since it is added to the side, complete reproduction of the input signal is possible.

[Brief description of the drawings]

FIG. 1 is a block diagram showing the configuration of an image processing apparatus according to a first embodiment of the present invention, FIGS. 2A to 2G show conventional LOT transformation and inverse transformation, and FIG. FIG. 4 is a diagram showing an orthogonal transformation method, FIG. 4 is a block diagram showing an internal configuration of an encoder in FIG. 1, FIG. 5 is a block diagram showing an internal configuration of a decoder in FIG. 1, and FIG. FIGS. 7 (a) and 7 (b) are explanatory diagrams for explaining an orthogonal transformation method, and FIGS. 7 (a) and 7 (b) are block diagrams showing the configuration of the main part of an image processing apparatus according to a second embodiment of the present invention.
It is a figure showing the case where dimension LOT transformation is performed. 54a, 54b, 54c, 54d: encoder, 58a, 58b, 58c, 58d: decoder, 54a-2, 54a-7, 58a-5, 54a-C, 54a-I: buffer, 54a -3,54a-D ... LOT converter, 54a-6,58a-
4,54a-H LOT inverse converter.

Claims (2)

(57) [Claims] 1. An orthogonal transformation method for sequentially extracting a finite-length input signal to generate a plurality of signal blocks, performing an LOT transform on each of the signal blocks, and encoding the input signal, wherein: The orthogonal transformation method, wherein the input signal is set in advance so that the start point and the end point of the signal are continuous, and the extraction is performed. 2. A plurality of signal blocks are generated by sequentially cutting out a finite-length input signal, and a plurality of transform coefficients are obtained by performing an LOT transform on each of the signal blocks. In the orthogonal transformation method of performing a LOT inverse transform to generate a plurality of local reproduction signals and adding the respective local reproduction signals to reproduce the input signal, the starting point of the first local reproduction signal among the respective local reproduction signals The orthogonal transformation method characterized in that the reproduction is performed by adding the side to the end point side of the last local reproduction signal.