KR100214378B1 - Adaptive quantizing method needless of transmitting additional information - Google Patents

Abstract

The present invention relates to an adaptive quantization method that achieves the effect of having an infinite number of representative vectors with a finite number of representative vectors, and does not need to transmit overhead information to a receiver.

Collecting all kinds of image data that can be generated and subdividing them into P × Q-sized vectors and classifying them into cell representative vector groups, and using the obtained cell vector groups, as many as the number of labels corresponding to the transmission bit rate Classifying the cell vectors to obtain a wide representative vector, and after obtaining the wide representative vector, assigning a local representative vector by a predetermined number using the cell representative vector that is closest to each wide representative vector (as the Eulidean Distance). It consists of steps.

As a result, not only the finite number of representative vectors and the infinite number of representative vectors are obtained, but the additional information is not transmitted to the receiver.

Description

Adaptive Quantization Method without Additional Information Transmission

1 is an explanatory diagram showing a propagation direction of an inter-frame block in moving image processing.

2 is a block diagram showing a wide area, a region, and a cell representative vector group.

3 is a diagram illustrating a process of generating a learning vector, a cell, a region, and a broad representative vector.

4 is a diagram illustrating a process of synchronizing a global representative vector label map at an encoder and a decoder in a video compression process.

5 is a diagram illustrating a method for setting transmission bits of a wide representative vector label and a local representative vector label, and is divided into one bit of a search mode bit and an N-1 bit of transmission code bits when N bits are designated by the bit rate. Figure.

6 shows an update of a global representative vector.

FIG. 7 is a block diagram of selecting a nearest wide representative vector from the first image and transmitting the label to the decoder before transmitting the first frame to improve adaptability in the video compression process.

8 shows the relative relationship between post and cell vectors and the post and updated global vectors.

9 is a predetermined distance map between a post and cell representative vector and a distance map between a post and a wide representative vector updated during execution.

The present invention relates to an adaptive quantization method that does not require additional information transmission. The present invention relates to an adaptive quantization method that obtains an effect of having an infinite number of representative vectors with a finite number of representative vectors, and does not require transmission of additional information to a receiver. It is about.

In particular, the present invention is an adaptive coding technique essential in video telephony, video conferencing systems, voice mailboxes and all systems requiring data compression processing, and also in ultra-low speed data compression processing such as MPEG-4.

It is very important to compress the data to minimize the amount of data to be sent when transmitting video or audio. In particular, such as in the case of video telephony or video conferencing system, it is essential for ultra-low speed transmission method. Until now, transform coding using Discrete Cosine Transform has been widely used (eg MPEG-1, MPEG-2, International Standards). Since data is processed in scalar units, there is a limit to bit rate reduction.

Therefore, there is a vector quantization technique that processes data in vector units to obtain a high compression ratio, and transmits only the label (address of the processed data), and is suitable for data processing for ultra-low speed transmission. It has been known as.

The general vector quantization method consists of a learning phase and an execution phase. First, in the learning stage, various kinds of possible data are classified (hereinafter referred to as a data group), and the classified training data are separated into a set of P × Q size vectors designated by the user.

From the set of vectors, the finite Nr representative vectors are extracted using an iterative clustering algorithm (e.g., K-means algorithm), and the representative vectors are codebooks of the data group. codebook).

In the execution step, the input data is separated into a set of vectors of size P × Q, a representative vector closest to each vector is selected, and only its label is transmitted to the decoder.

In the decoder stage, the corresponding input data is reproduced by replacing the representative vector in the codebook (the encoder and the decoder owns the same code; a kind of promised code) in the codebook that owns the received label in advance.

The vector quantization method divides the data into vector units and thus has an advantage over the scalar quantization method in terms of data compression efficiency. However, the vector quantization method has the following disadvantages.

1. During learning, the representative vector closest to the input data is replaced with the input data in the fixed codebook, so distortion of the data to be sent and the transmitted data occurs.

Accurate reconstruction of data minimizes distortion, and theoretically, if infinite representative vectors exist, distortion can be eliminated, but a large number of bits are required for labeling of nearly infinite representative vectors. (For example, 8 bits can represent 256 labels.) Contrary to data compression.

In fact, even if there is a representative vector close to infinity, only a few representative vectors are used to process each input data, and the majority of the representative vectors are not used. (There are many types of representative vectors to process all the different characteristics of the image, but when a specific image is input, the representative vector is not used to express other characteristics. For example, a simple image with few edges can be processed. Representative vectors that can process complex images are not used at all). In order to obtain the data compression efficiency, the number of finite or Nr representative vectors must be maintained and a vector having a large dimension must be used. Where Nr ∞.

2. Since the representative vector is extracted by an iterative classification method from a finite group of training data, it cannot generate a codebook suitable for all input data.

Since the codebook determined at the learning stage cannot be arbitrarily changed, it cannot adaptively adapt to any input data (since the promised codebook must exist at the same time as the encoder unit and the decoder unit). In order to adapt adaptively to the data to be transmitted, when processing each data, it is necessary to reset the appropriate representative vector and transmit the codebook to the receiver, and then transmit the label of the representative vector closest to the data to be transmitted. There is a way. Let's call this general Adaptive Vector Quantization (Goldberg, M., Boucher, PR and Shlien, S., (1986) Image Compression Adaptive Vector Quantization, Ieee Trans, Commun., Vol COM-34) , see pp. 180-187), but since the representative vector must always be transmitted, the amount of overhead increases significantly. In order to increase the compression efficiency, the amount of additional information should be reduced by reducing the number of representative vectors.

At this time, the distortion with the data to be expressed increases in proportion (the number of representative vectors). Exact data recovery using only a small number of representative vectors is difficult.

SUMMARY OF THE INVENTION The present invention has been made to overcome the above-mentioned problems, and an object of the present invention is to provide an adaptive vector quantization technique in a direction that reduces the aforementioned two disadvantages, and provides two conditions for an ideal adaptive vector quantization technique. the ideal adaptive vector quantization methodology, which satisfies

1. The effect of having infinite representative vectors with finite representative vectors is obtained.

2. No overhead is transmitted to the receiver.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

The present invention satisfies the following as two conditions for the ideal adaptive vector quantization methodology.

1. The effect of having infinite representative vectors with finite representative vectors is obtained.

2. No overhead is transmitted to the receiver.

The same amount of data is present continuously (a certain amount of data is called a data frame; for example, a video frame, an audio frame), the correlation between frames is relatively high, and each frame When it is possible to predict the information flow in the information (Information Flow) it is possible to implement an adaptive vector quantizer that satisfies the two conditions specified above. For example, as shown in FIG. 1, when processing video frame data, each frame is divided into several fixed blocks, and any block in the current frame has the highest correlation with any block in the previous frame. In this case, when the position of the block in the frame is the same between frames as shown in (a) of FIG. 1 or motion compensation (MC) as shown in (b) of FIG. If it is possible to predict where the position of the current frame is propagated in the previous frame, an adaptive vector quantizer that satisfies the above condition can be implemented.

In order to satisfy the first condition of the ideal adaptive vector quantization method (the effect of having infinite representative vectors with infinite representative vectors), the representative vectors are represented by a global representative vector group (GRV) and a local representative vector group. (Local Representative Vector: LRV) and Cell Representative Vector (CRV) are divided as shown in FIG.

The cell representative vector group collects all kinds of image data that can be generated, designates it as learning data, divides it into a P SQ size vector, and sets it through a classification process.

As the classification method, the conventionally used vector quantization method and the like described above are used. That is, in the present invention, as a learning method for creating the cell representative vector group, a general vector quantization method consisting of a learning step and an execution step as described above is used.

Using the obtained cell vector group, a wide-area representative vector is obtained by classifying cell vectors by the number of labels corresponding to the transmission bit rate (the transmission bit rate and the number of labels are directly proportional to the number of fixed vectors).

After obtaining the global representative vector, the local representative vector is designated by a predetermined number using the cell representative vector closest to each of the global representative vectors (as the Eulidean Distance). (Here it is equal to the number of representative vector vectors). The number of cell vectors can create limits to the system's ability to process (speed of processing or storage space).

Therefore, Condition 1 can be satisfied by adjusting the finite wide-area vector and the local-representative vector in a space having a nearly infinite number of representative vectors. This is shown in FIG.

In the initial setting of the cell and the regional wide area representative vector, all the video information of the video frame is collected and initialized. That is, the representative vectors are initialized using the collected image information data as learning data of a learning process for extracting the representative vector. In the initial setting of the representative vectors, vector quantization is used in an INTRA frame mode, and motion compensation is used in an INTER mode. If motion compensation fails, the same vector quantization method as in the intra frame mode is used. In order to satisfy the second condition of the ideal adaptive vector quantizer, that is, no additional information is sent, the same global representative vector label map (GRVLM) is generated and synchronized to the encoder and the decoder. A wide representative vector label map in the video data compression process is shown in FIG.

The label map is formed by storing a label of a wide representative vector in a macro block that is coded and transmitted at the beginning of the compression process, and synchronization uses motion compensation coefficients.

That is, the label of the global representative vector label map in the macro block is replaced using the motion compensation coefficient that is essentially transmitted from the encoder to the decoder.

The transmission of the signal from the encoder to the decoder is by the label bitmap shown in FIG. Assuming that N bits are required to display one label to maintain a predetermined transmission bit rate, a search mode bit (SMB) of one bit is used to represent two representative vectors of a wide area and a region. Assign N-1 bits to the label representation.

That is, 2 ^(N-1) wide representative vectors each have 2 ^(N-1) local representative vectors.

SMB is designated to transmit the label of the wide representative vector, SMB is designated as 1 to transmit the label of the local representative vector, and the corresponding label of N-1 bits is transmitted.

The adaptation of the vector quantizer in the moving picture compression process is performed by the wide representative vector label map shown in FIG. 4 and the transmission bit setting method of the label shown in FIG. 5, and the update of the wide representative vector shown in FIG. It is obtained by repeating the (Update) method.

In other words, the adaptive method of the vector quantization method in the moving picture data compression process will be described as follows.

1. Synchronization of Wide Representative Vector Label Map

The motion compensation coefficients (MC Coefficients) are obtained by comparing the previously input, coded, transmitted, and reproduced images that are currently input from the encoder, and using the motion compensation coefficients, Change the label map for the part (the moving part).

That is, the first image of the first video frame to be transmitted is encoded in an intra mode that does not use motion information. Therefore, the global representative vectors that best represent the image are obtained from the cell vectors, and the first image is encoded to transmit the labels for the vectors.

In the intra mode, a general vector quantization method is applied, wherein the vector size of the macro block illustrated in FIG. 4 represents the size (or length) of P × Q. Here, the label map consists of a 2D plane according to the positional relationship of the macro blocks, and the lammel map has a label value of a wide vector instead of image information.

In addition, after the first image is transmitted, the second image is encoded in an INTER mode in which motion is estimated from the reconstructed previous image. That is, the previous macro information is restored by using the transmitted motion information to restore the current macro block. In this case, when the image quality of the image reconstructed by the motion compensation is deteriorated, a wide area or region representative vector is selected and transmitted instead of the motion information.

Therefore, the vector label value of the label map is readjusted using the motion information even for the label map constructed in the first image encoding process. The decoder also uses the transmitted MC coefficients to change the label map to synchronize with the encoder. (4th degree)

2. Determining and Sending Representative Vectors

When it is necessary to code and transmit an inconsistency caused by a lack of compensation (or a new object added to the entire image) by comparing the current input video with the motion-compensated playback video (step 2-1), the part to be newly coded Find the label of the wide-area representative vector nearest to. (Step 2-2) If the label value is different from the label of the corresponding wide area representative vector and the label map (GRV Label Map) of the corresponding position (wide area representative vector label of the previous frame), and the label value is different, as shown in (b) of FIG. As SMB is set to 0, the label value of the wide representative vector is transmitted to the decoder. If the values of the labels are the same, in order to represent the transmitted data more accurately, as shown in (c) of FIG. 5, SMB is designated as 1, and 2 ^(N-1) regions closest to the position of the wide representative vector corresponding to the label value are ^shown. The local representative vector is selected among the representative vectors closest to the part to be newly coded and the label is transmitted. By using this method, the data ^{(2) (N-1) has} the option of (number of GRV Labels × number of LRV Labels) to achieve accurate restoration of data.

3. Global Representative Vector Update

The update interval is performed for each image frame or when there is a change in a certain number of global representative vectors.

(Step 3-1) When an area representative vector belonging to a specific wide area representative vector is designated and transmitted many times as shown in FIG. 6 in the encoder and the decoder (step 3-2), all the coded area representative vectors are collected. Find the centroid and designate it as the new global representative vector. (Step 3-3) The closest N-1 cell representative vectors are selected based on the new global representative vectors and designated as local representative vectors.

As shown in (b) of FIG. 5, SMB is set to 0 and the value of the wide representative vector is subtracted from the vector value to be sent to correctly send special objects when the label value of the wide representative vector is transmitted to the decoder. You can also think of how to code and transmit the difference (frame difference).

As mentioned above, even if the vector quantization method has a codebook with a very large number of representative vectors, only a few representative vectors are required to code and transmit a specific image when it is input. For example, when processing a very simple image, a representative vector representing a high energy portion is not used.)

In the compression process, which is highly correlated with the initial screen, such as moving image processing, and the image connected later is generally the same complexity as the initial screen, the wide-area representative vector closest to the initial screen is transferred from the cell representative vector group before coding and transmitting the initial screen. After obtaining and setting, a wide-area representative vector initialization method for transmitting only the label to the decoder stage is proposed.

By doing this, it is possible to adapt to the data to be transmitted from the beginning, and by transmitting only a label, the amount of bits transmitted is very small. This is shown in FIG.

In the present invention, the wide-area representative vector is a vector quantization method characterized by moving in order to cope with the input image adaptively. At this time, a regional representative vector should be set from the neighboring cell representative vector group for each newly updated wide area representative vector.

It is common to simply obtain the absolute distance between all cell vectors and the updated global representative vector to obtain the nearest constant number (two ^(N-1) regions) in the present invention, but the computational amount is so large that the real time Processing is impossible. Therefore, the present invention proposes a method of updating a local representative vector in a relative distance relationship by having three or more posts in the space of the cell representative vector group (which can be obtained by knowing the dimensions and sizes of the cell vectors). do.

As shown in FIG. 8, three or more posts are placed in the space occupied by the cell representative vector group, and the distance between each cell representative vector and the post is processed in the learning stage by creating a distance map as shown in FIG. Make a table. This is handled in the learning phase, so a lookup map can be owned by both the encoder and the decoder at the same time.

In real time execution of the vector quantization method as in the video compression process, the distance between the updated representative vector and the post is obtained as shown in FIG. 8, and the distance map is shown in FIG. 9 (b). Using the interval map of FIG. 9 (b) and the interval map of FIG. 9 (a), a representative number of cell representative vectors ⁽ K 2 ^(N-1) ), which is relatively close to the wide-area representative vector whose distance is updated Then, the distance between the K cell vectors and the updated wide area vector is obtained, and 2 ^(N-1) local representative vectors are set.

Using the above method, we can reduce the amount of computation and enable real-time processing by using the wide vector and post and the relative distance between each cell vector and post when setting up the local representative vector.

Claims (17)

Collecting all kinds of image data that can be generated, subdividing them into P × Q size vectors, and then classifying a cell representative vector group through a classification process, and using the obtained cell vector group, as many as the number of labels corresponding to the transmission bit rate Classifying the cell vectors to obtain a wide representative vector, and after obtaining the wide representative vector, assigning a local representative vector by a predetermined number using the cell representative vector that is closest to each wide representative vector (as the Eulidean Distance). Adaptive quantization method that does not require the transmission of additional information, characterized in that consisting of a step. The method of claim 1, wherein the setting of the cell vector group is performed by using a learning method used in a general vector quantization method. The method of claim 1, wherein the number of global representative vectors and the number of local representative vectors are the same. 2. The method of claim 1, wherein the number of cell vectors can make the number of representative vectors near infinite to the limit that the system can process. 5. The method of claim 4, wherein in the step of obtaining the global representative vector, the transmission bit rate and the number of labels are directly proportional to each other. After collecting all kinds of image data that can be generated and segmenting them into P × Q size vectors, a cell representative vector group is established through the classification process, and using the obtained cell vector group, the cell vectors as many as the number of labels corresponding to the transmission bit rate are obtained. Classifying a to obtain a global representative vector, obtaining a global representative vector, and then designating a local representative vector by a predetermined number using a cell representative vector closest to each global representative vector (as the Eulidean distance), The motion compensation coefficient is obtained by comparing the video currently input from the encoder with the previously input, coded, transmitted, and replayed video, and using the motion compensation coefficient, The label map is changed, and the decoder also changes the label map using the transmitted motion compensation coefficients. Adaptive quantization method that does not require the additional information transmitted, characterized in that consisting of a synchronization phase of the vector representative of wide-area label map for an encoder and synchronization. The part to be newly coded when it is necessary to code and transmit an inconsistency caused by a lack of compensation (or a new object is added to the previous image) by comparing the current input image with the playback image with motion compensation. Obtaining a label of the widest representative vector closest to each other, and comparing and determining a label of the obtained global representative vector and a label of a corresponding position in the label map (a wide representative vector label of the previous frame). Adaptive quantization method that does not require additional information transmission, characterized in that it further comprises. The label value of the global representative vector according to claim 7, wherein the label of the obtained wide representative vector and the label of the corresponding position in the label map (wide representative vector label of the previous frame) are compared with each other. Adaptive quantization method that does not require additional information transmission, characterized in that for transmitting to the decoder. The method according to claim 8, wherein when SMB is set to 0 and the label value of the wide representative vector is transmitted to the decoder, the difference (frame) is obtained by subtracting the vector value to be sent and the wide vector being transmitted in order to correctly send special objects. Adaptive quantization method that does not require the transmission of additional information, characterized in that for coding the difference; Frame Difference. The method of claim 6, wherein when the labels of the obtained wide representative vectors and the labels of the corresponding positions in the label map (wide representative vector labels of the previous frame) are compared with each other, and the values of the labels are the same, the data to be transmitted is more accurately represented. In order to specify the SMB as 1, select the region representative vector that is closest to the newly coded part of the 2 ^(N-1) region representative vectors that are closest to the position of the global representative vector corresponding to the label value, and transmit the label. An adaptive quantization method requiring no additional information transmission. 7. The method of claim 6, wherein the global representative vector is updated for each image frame or when there is a change in a certain number of global representative vectors, wherein the updating step includes a local representative vector belonging to a specific global representative vector in the encoder and the decoder. When multiple times specified and sent, gathering all the encoded local representative vectors to obtain their centroids and assigning them as new global representative vectors, and the 2 ^(N-1) cells closest to the new global representative vectors An adaptive quantization method requiring no additional information, characterized by selecting a representative vector and designating it as a local representative vector. The method of claim 6, wherein the label of the global representative vector in the macroblock that is coded and transmitted at the beginning of the compression process is stored to form a global representative vector label map, and the motion compensation coefficient that is essentially transmitted from the encoder to the decoder is used. And synchronous quantization of the wide-area representative vector label map in the macroblock by the encoder and the decoder. The label representation of claim 6, wherein assuming that N bits are required to display one label to maintain a predetermined transmission bit rate, one bit of a search mode bit is used to represent two representative vectors of a wide area and a region. Adaptive quantization method without the need for additional information transmission, characterized in that each of the ^{N (1)} bits, so that each of the 2 ^(N-1) wide representation vector has 2 ^(N-1) local representation vector. The method of claim 6, wherein the SMB is designated as 0 to transmit the label of the wide representative vector, and the SMB is designated as 1 to transmit the label of the local representative vector, and the additional information is transmitted. Adaptive quantization method without. 7. The adaptive quantization method as claimed in claim 6, wherein the setting of the cell, region, and wide representation vector is performed by collecting all of the intra-frame images and performing vector quantization in an intra frame mode. The method according to claim 6, wherein a wide region representative vector closest to the initial screen is transmitted before coding and transmitting the initial screen in a compression process having a very high and connected image, which is largely independent of the initial screen in the moving picture, and has the same complexity as the initial screen. An adaptive quantization method requiring no additional information, characterized in that only the label is transmitted to the decoder stage after obtaining and setting from the representative vector group. By placing three or more posts in the space occupied by the cell representative vector group and creating a distance map of the distance between each cell representative vector and the post, the table is referred to as a reference table in the learning step. In real-time execution, obtaining a distance between the updated representative vector and the post and referencing the spacing map, and using the spacing maps, a constant number close to the relatively wide area representative vector updated using the spacing map (K = K) ^Obtaining a cell representative vector of ^(N-1) , and setting two ^(N-1) local representative vectors by calculating a distance between the K cell vectors and the updated wide area vector. Adaptive quantization method that does not require information transmission.