KR101241508B1 - Methods of motion vector estimation using classfier and apparatuses for using the same - Google Patents

Abstract

Disclosed are a motion vector estimation method using a classifier and an apparatus using the method. A motion vector estimation method using a classifier includes one of a first block that performs motion prediction and a second block that does not perform motion prediction based on the calculated feature vector. The method may include classifying and generating a prediction block by performing motion prediction on the first block, and generating a prediction block without performing motion prediction on the second block. Therefore, the complexity of the computation can be reduced by determining whether the block performs the motion prediction in advance through the classifier.

Description

METHODS OF MOTION VECTOR ESTIMATION USING CLASSFIER AND APPARATUSES FOR USING THE SAME}

The present invention relates to a motion vector estimation method using a classifier and an apparatus using the method, and more particularly, to a method and apparatus for estimating a motion vector.

Recently, as the information media handled by communication networks ranging from telephones to facsimile, data, and images become larger, the interest and demand for high-speed multimedia data communication is increasing. Among the multimedia data communication fields that are currently receiving the most attention is the transmission and reception of multimedia data through a wireless network. As the capacity of the wireless data transmission / reception channel is greatly improved, multimedia data having a large data size can be transmitted and received in real time through the wireless communication network channel.

Because video data is quite large, there are difficulties in storing and transmitting it. In addition, since video information must be appropriately processed in accordance with the requirements of the application, a system for processing video is also required. In order to solve these problems, video compression algorithms, which are techniques for converting video data into small information, have been actively studied. Representative international standards related to video compression studied to date include the MPEG series of ISO / IEC and the H.26x series of ITU-T.

Video data has three redundancies due to its characteristics. Such redundancy includes temporal, spatial, and statistical redundancy. Temporal redundancy means redundancy that exists between successive frames. The brightness values of pixels that exist at the same position in successive frames are statistically similarly correlated, also called "inter-frame redundancy." . In addition, spatial redundancy means redundancy existing in a frame. The brightness value of one pixel has a similar correlation with the brightness value of neighboring pixels, which is also referred to as "inter-frame redundancy". Finally, statistical redundancy means redundancy between coded data, redundancy between pixels and redundancy in encoding.

In the currently known moving picture coding standard, by eliminating these three redundancies, a huge amount of moving pictures can be compressed into smaller data. For example, transform and quantization processes, such as the Discrete Cosine Transform (DCT), are used for spatial redundancy, entropy coding, such as variable field coding, for statistical redundancy, and motion estimation and compensation for temporal redundancy. It is adopted and used in international standard of video codec.

Among these, motion estimation and compensation methods for removing temporal redundancy can be broadly classified into a pixel recursive algorithm (PRA) and a block matching algorithm (BMA). Among these, most video codec technologies use block matching algorithms in consideration of the regularity of data flow, computational complexity, and hardware implementation. The block matching algorithm divides a screen into 16x16 macroblocks (macroblocks) or smaller search blocks of any size, and finds the matching blocks within a predetermined search area of the reference frame in units of blocks. Point to. The value indicating the position of the matching block selected based on the block matching algorithm is a motion vector (MV). In video encoding, only the difference value and the motion vector between the matching block and the current block are encoded to remove redundancy of data.

The block matching algorithm may generally use a full search algorithm (FSA). The global search algorithm refers to a matching error, for example, a sum of absolute differences between all pixels constituting the search block of the current block and the reference frame, using all coordinates in the search area as candidate points (CP). Absolute Difference (SAD) is used.

Most video coding standards use the global search algorithm using the SAD, which has the advantage of finding the optimal motion vector because the best matching candidate block is selected as the matching block. However, the global search algorithm has a large number of search points and a large amount of computation, and a complicated system configuration for implementing the global search algorithm is difficult to apply to an application for real-time encoding.

Research on high-speed algorithms to reduce the computational complexity of various global search algorithms has been conducted. Successive Elimination Algorithm (SEA) using block sums of reference and candidate blocks, and three steps of selectively searching without searching all points in the search area. Match scan using Three Step Algorithm, New Three Step Algorithm, Four-step Search Algorithm, Diamond Search Algorithm, and Image Complexity As a fast algorithm that reduces the amount of computation by rearranging the order, Motion Vector Field Adaptive Search Technique (MVFAST), Predictive Motion Vector Field Adaptive Search Technique (PMVFAST), etc. There is this.

Nevertheless, these algorithms failed to find an optimal solution and the fast motion prediction problem remains an important research topic. In particular, the binary block motion estimation method greatly reduces the complexity required for conventional motion prediction by mapping one frame of several numerical pixels to one frame of binary numerical pixels. After converting the current frame and the reference frame into binary form, the matching error is reduced by using exclusive-OR (XOR) operation to reduce the complexity.

The binary block motion estimation method is also less satisfactory than the block matching algorithm, and the binary block motion estimation method has a disadvantage in that the image quality is significantly lower than the global search method using the SAD. .

Accordingly, a first object of the present invention is to provide a method for determining whether to estimate a motion vector of a block using a classifier.

It is also a second object of the present invention to provide a classifier for determining whether to estimate a motion vector of a block.

It is also a third object of the present invention to provide an encoder for determining whether to estimate a motion vector of a block.

Motion vector estimation method using a classifier according to an aspect of the present invention for achieving the first object of the present invention described above calculates a feature vector for determining whether to predict motion in a block and based on the calculated feature vector Classifying the signal into one of a first block that performs motion prediction and a second block that does not perform motion prediction and performing motion prediction on the first block to generate a prediction block, and predicting motion on the second block. It may include generating a prediction block without performing the operation. The motion vector estimation method using the classifier may further include performing 1-bit transform on the reference frame and the current frame. Generating a prediction block by performing motion prediction on the first block, and generating a prediction block without performing motion prediction on the second block may be performed on the reference frame and the current frame on which the 1-bit conversion is performed. The motion prediction may be performed based on a binary block motion estimation method. The motion vector estimation method using the classifier may further include determining a feature vector to be used for classification and performing preprocessing to calculate the feature vector. Calculating a feature vector for determining whether to predict motion in the block and classifying the block into one of a first block for performing motion prediction and a second block for not performing motion prediction based on the calculated feature vector; Based on a conditional probability value calculated as a result of learning using the image of, if the block has a predetermined feature vector, it may be determined whether to calculate a motion vector by performing motion prediction on the block. Computing a feature vector that determines whether to predict motion in the block, and classifying the block into one of a first block that performs motion prediction and a second block that does not perform motion prediction based on the calculated feature vector. (Bayesian) classification method, Perceptron classification method, SVM (Support Vector Machine) can be used. The feature vector includes a quantization coefficient for a block to be encoded, a frame size, a block size, a predicted motion vector size, a number of encoded neighboring blocks in a picture, a number of skip encoded neighboring blocks, and the 1-bit transform. It may be at least one of the number of 1's output through and the difference between the pixel value passed through the band pass filter and the original pixel value during binary conversion.

In addition, the classifier for determining whether to estimate the motion vector of the block according to an aspect of the present invention for achieving the above-described second object of the present invention calculates a feature vector for determining whether to calculate the motion prediction vector of the block And a block classification unit classifying whether to perform motion prediction on the block based on the feature vector calculator and the feature vector calculated from the feature vector calculator. The classifier for determining whether to estimate the motion vector vector of the block may further include a 1-bit converter configured to perform 1-bit conversion of the current frame and the reference frame by performing a predetermined filtering. The classifier for determining whether to estimate the motion vector of the block may further include a feature vector determiner that is used to classify the block and a preprocessor that performs preprocessing to calculate the feature vector. The block classifier may determine whether to perform motion prediction on the block when the block has a predetermined feature vector based on a conditional probability value calculated as a result of learning using a plurality of images.

In addition, the encoder according to an aspect of the present invention for achieving the above-described third object of the present invention is a classifier for classifying whether to calculate a motion vector in a block and a motion prediction classified by the classifier to calculate a motion vector. It may include a motion compensation unit for generating a prediction block by performing motion prediction on the performance block. The classifier determines whether to perform motion prediction in the block based on a feature vector calculator for calculating a feature vector for determining whether to calculate a motion prediction vector of the block and the feature vector calculated from the feature vector calculator. The block classification unit may be classified. The block classifier may determine whether to perform motion prediction on the block when the block has a predetermined feature vector based on a conditional probability value calculated as a result of learning using a plurality of images. The classifier may further include a 1-bit converter which performs 1-bit conversion of the current frame and the reference frame by performing a predetermined filtering.

As described above, according to the motion vector estimation method using the classifier and the apparatus using the method according to the embodiment of the present invention, the complexity of calculation can be reduced by determining whether the block performs the motion prediction through the classifier.

1 is a flowchart illustrating a motion vector estimation method using a classifier according to an embodiment of the present invention.
2 is a flowchart illustrating a method of training a classifier in a motion vector estimation method using a classifier according to an embodiment of the present invention.
3 is a flowchart schematically illustrating a motion vector estimation method using a classifier according to an embodiment of the present invention.
4 is a block diagram illustrating an apparatus for estimating a motion vector using a classifier according to an embodiment of the present invention.
5 is a conceptual diagram illustrating an encoder including a motion vector estimating apparatus using a classifier according to an embodiment of the present invention.

While the invention is susceptible to various modifications and alternative forms, specific embodiments thereof are shown by way of example in the drawings and will herein be described in detail. It should be understood, however, that the invention is not intended to be limited to the particular embodiments, but includes all modifications, equivalents, and alternatives falling within the spirit and scope of the invention. Like reference numerals are used for like elements in describing each drawing.

The terms first, second, etc. may be used to describe various components, but the components should not be limited by the terms. The terms are used only for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, the first component may be referred to as a second component, and similarly, the second component may also be referred to as a first component. And / or < / RTI > includes any combination of a plurality of related listed items or any of a plurality of related listed items.

It is to be understood that when an element is referred to as being "connected" or "connected" to another element, it may be directly connected or connected to the other element, . On the other hand, when a component is referred to as being "directly connected" or "directly connected" to another component, it should be understood that there is no other component in between.

The terminology used herein is for the purpose of describing particular example embodiments only and is not intended to be limiting of the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this application, the terms "comprise" or "have" are intended to indicate that there is a feature, number, step, operation, component, part, or combination thereof described in the specification, and one or more other features. It is to be understood that the present invention does not exclude the possibility of the presence or the addition of numbers, steps, operations, components, components, or a combination thereof.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. Hereinafter, the same reference numerals are used for the same components in the drawings, and duplicate descriptions of the same components are omitted.

Hereinafter, a feature vector used in a motion vector estimation method using a classifier according to an embodiment of the present invention is a vector representing a feature of a block for determining a motion estimation method of a predetermined block. Can be divided into a set of dogs. For example, the quantization coefficient, the screen size, the block size, the predicted motion vector size, the number of coded neighboring blocks, the number of skip coded neighboring blocks, and the binary block motion for the block to be encoded as the feature vector. Various values that can classify the characteristics of the block, such as the number of 1s obtained through the estimation and the difference between the pixel that passed through the band pass filter and the original pixel during binary conversion, are used as the feature vectors of the block.

A motion vector estimation method using a classifier according to an embodiment of the present invention may be performed based on a binary block motion estimation method. The binary block motion estimation method is a motion estimation method based on a 1 bit transform method and refers to a method of converting an 8-bit pixel value into 1 bit.

1 is a flowchart illustrating a motion vector estimation method using a classifier according to an embodiment of the present invention.

Referring to FIG. 1, one-bit conversion is performed on a reference frame (step S100).

According to the motion vector estimation method using the classifier according to an embodiment of the present invention, the one-bit conversion method to be described below is one example of the one-bit conversion method, and other one-bit conversion methods may also be used as long as they do not depart from the essence of the present invention. have.

When the 1-bit transform is used for motion estimation, the computational complexity is reduced compared to the block matching based motion estimation method.

One-bit conversion may use a filter kernel to convert pixels with various pixel values into pixels with one bit values.

Equation 1 below is a determinant of an exemplary filter kernel for performing a 1-bit transform, and has the property of a band pass filter.

Determinants different from Equation 1, which are filtering matrices for 1-bit conversion, can also be used and such filter kernels are also within the scope of the present invention.

The pixel value of the reference frame filtered using Equation 1 above may be expressed as 0 or 1 based on Equation 2 below after comparing with the pixel value of the original reference frame.

은 필터링을 수행하지 않은 참조 프레임의 픽셀값을 나타내고

는 필터링을 수행한 참조 프레임의 픽셀값을 나타낸다.

Denotes the pixel value of the reference frame without filtering

Denotes a pixel value of the reference frame on which the filtering is performed.

One-bit conversion is performed on the current frame (step S110).

One bit conversion for the current frame may also be one bit converted in the same manner as described in step S100 described above.

The feature vector is determined (step S120).

As described above, the feature vector may be used by various variables representing the characteristics of the block, and one of the features may be determined as a feature vector for classifying the block. A motion vector estimation method using a classifier according to an embodiment of the present invention may determine whether to perform motion estimation based on a feature vector of a block.

When classifying blocks based on whether motion estimation is to be performed or not, an effective feature vector may vary according to the characteristics of the image, and thus the feature vector may be adaptively selected based on the characteristics of the image.

However, according to the motion vector estimation method using the classifier according to an embodiment of the present invention, the feature vector may be used as a fixed variable value, and when the feature vector is a fixed variable, determining what to do with the feature vector S120 May not be performed.

The preprocessing for calculating the feature vector is performed (step S130).

Preprocessing can be performed if necessary to produce a feature vector clearly. For example, when performing binary block motion estimation, the number of 1s of pixel values of a block converted by 1-bit transform may be used as a feature vector. Therefore, the preprocessing step of calculating the number of 1s included in the binary block may be performed. If this preprocessing step is unnecessary and the feature vector can be calculated immediately, step S130 may not be performed.

Also, the preprocessor may classify the feature vectors into predetermined categories according to the feature vectors. For example, in H.264 / AVC, the quantization parameter has a value from 0 to 51. When the quantization parameter is a feature vector, the quantization parameter is classified into four predetermined intervals to reduce the computational complexity. Can be used to classify feature vectors. That is, when the quantization parameter values from 0 to 12 are included, the first feature vector 1 and the quantization parameter values from 13 to 25 may be classified as the second feature vector to reduce complexity when calculating probability values. If such classification is unnecessary in a specific vector, step S130 may not be performed.

The feature vector is calculated (step S140).

4, 4

8, 8

4, 8

8, 8

16, 16

8, 16

16...32

32...64

64 등 다양한 사이즈의 블록이 될 수 있다. 즉, 움직임 벡터를 산출하는데 사용될 수 있는 다양한 크기의 블록은 특징 벡터를 산출할 때 또한 이용될 수 있다.A feature vector may be calculated from blocks included in the current frame. The block from which the feature vector is calculated is 4

4, 4

8, 8

4, 8

8, 8

16, 16

8, 16

16 ... 32

32 ... 64

It can be a block of various sizes such as 64. That is, blocks of various sizes that can be used to calculate the motion vector can also be used when calculating the feature vector.

When determining a block to calculate a feature vector, a block of various sizes for calculating a motion vector may be used in one image based on characteristics of an image such as an edge.

The feature vector is a vector representing a feature of a block for determining a motion estimation method of a predetermined block, as described above. For example, a quantization parameter, a screen size, a block size, a predicted motion for a block to be encoded. Characteristic of blocks such as vector size, number of coded neighboring blocks, skip coded neighboring blocks, number of 1s through binary block motion estimation, and difference between original pixel and bandpass filter during binary conversion Various values that can be classified into each may be used as the feature vector of the block.

The block is classified according to the feature vector, and motion prediction is performed on the motion prediction target block (step S150).

Blocks may be classified using various classification methods based on feature vectors.

Blocks may be classified into blocks that do not perform motion prediction and blocks that perform motion prediction.

Equation 3 below shows a Bayes classification method in which a block is classified into a block that does not perform motion prediction and a block that performs motion prediction in a motion prediction vector estimation method using a classifier according to an embodiment of the present invention. It is shown.

As a classifier used in the motion prediction vector estimation method using the classifier according to an embodiment of the present invention, at least one of various classifiers such as a Bayesian classifier, a Perceptron classifier, and a support vector machine (SVM) may be used. And classifying blocks using such classifiers are also within the scope of the present invention.

는 특징 벡터

가 주어진 경우

분류가 선택될 확률을 나타낸다.

Is a feature vector

Is given

Represents the probability that a classification will be selected.

Conditional probability values can be obtained through various training in the classifier. It is possible to determine whether the motion of the corresponding block is predicted based on the probability value obtained after learning the various images.

가 블록의 움직임 벡터 예측을 하지 않는 분류를 나타내고

가 블록의 움직임 추정을 하는 분류를 나타낼 경우, 특징 벡터

에 대해서

가 더 큰 경우, 특징 벡터

를 가진 블록에 대해서 움직임 벡터 예측을 수행하지 않고

가 작거나 같은 경우 특징 벡터

를 가진 블록에 대해서 움직임 벡터 예측을 수행할 수 있다. Like Equation 4 above

Represents a classification that does not predict the motion vector of the block

If the vector represents a classification that estimates the motion of the block, then the feature vector

about

Is greater, feature vector

Without performing motion vector prediction on blocks with

Is less than or equal to the feature vector

Motion vector prediction can be performed on a block with.

A block that does not perform motion prediction may estimate the motion vector value of the block based on, for example, the motion vector value of the neighboring block without performing the motion prediction. That is, other prediction methods that do not perform motion prediction may be used.

A block that performs motion prediction may calculate a motion vector value of the block based on various motion vector prediction methods. As a motion vector prediction method, various methods based on a binary block motion estimation method for predicting a motion vector based on a block on which a 1-bit transform method is performed may be used.

Equation 5 below represents one of the 1-bit transform based motion prediction methods.

는 XOR 연산을 나타내고

는 탐색 범위를 결정하며

와

는 이미지의 위치를 의미한다. 최소의 NNMP를 가지는 지점인

를 움직임 벡터로 설정할 수 있다.

과

은

의 범위를 가진다. NNMP로 8 비트의 메모리가 이용되는 경우 8개의 정합 오차, 32 비트의 메모리가 이용되는 경우 32개의 정합 오차가 산출될 수 있다.In Equation 5 above, the number of non-matching points (NNMP) is a number of non-matching points to calculate a matching error for two 1-bit image frames.

Represents an XOR operation

Determines the search range,

Wow

Means the position of the image. The point with the least NNMP

Can be set as a motion vector.

and

silver

Has a range of. Eight matching errors may be calculated when 8-bit memory is used as NNMP, and 32 matching errors may be calculated when 32-bit memory is used.

In addition to the above-described NNMP method among binary block motion estimation methods, various 1-bit transform based motion prediction methods may be used, and motion prediction methods using such 1 bit transform based motion prediction methods are also included in the scope of the present invention.

After the motion prediction is performed, a packet encoding the error block and the motion prediction vector may be generated. The blocks on which motion prediction is performed and the blocks on which motion prediction is not performed may be quantized and encoded using an encoding method such as an entropy encoding method.

2 is a flowchart illustrating a method of training a classifier in a motion vector estimation method using a classifier according to an embodiment of the present invention.

A feature vector is calculated in the block (step S200).

As described above, the feature vector includes a quantization coefficient, a screen size, a block size, a predicted motion vector size, a number of coded neighboring blocks, a number of skip coded neighboring blocks, and a binary block motion for a block to be encoded. It can be various kinds of values such as the number of 1s obtained through the estimation and the difference value between the pixel which passed the band pass filter and the original pixel during binary conversion.

Such a feature vector may be calculated in the block, and the block may be classified based on the calculated feature vector.

The calculated feature vector is classified according to a predetermined criterion (step S210).

For example, in H.264 / AVC, the quantization parameter has a value from 0 to 51. When the quantization parameter is a feature vector, the quantization parameter is classified into four predetermined intervals to reduce the computational complexity. Can be used to classify feature vectors. That is, when having a quantization parameter value of 0 to 12, and having a quantization parameter value of the first feature vector 1, 13 to 25 can be classified into a second feature vector to reduce the complexity when calculating the probability value.

In the case of step S210, if the feature vector can be classified into a small number as an optional step, it may not be performed.

The motion prediction is performed (step S220).

For learning, it is necessary to compare the result of performing the motion prediction with the result of not performing the predetermined feature vector value and determine whether to perform the motion prediction according to a predetermined criterion. In the step of performing the prediction, the block for performing the motion prediction is a binary block that is a variety of 1-bit transform-based motion prediction methods in addition to the above-described NNMP method and NNMP method based on the reference frame on which the 1-bit transform is performed and the block included in the current frame. Motion estimation can be performed using the motion estimation method.

Therefore, in the motion vector estimation method using the classifier according to an embodiment of the present invention, the motion prediction is performed on all blocks in the learning process to generate a block in which the motion prediction is performed, and as shown in step S230 below without performing the motion prediction. It should be compared with blocks that terminate early without performing motion prediction.

A prediction block is generated by a method other than motion prediction (step S230).

A prediction block is generated using a method of predicting a motion vector of a current block from intra motion prediction or a motion vector of a neighboring block without performing motion prediction.

It is determined whether to perform motion prediction (step S240).

Various criteria may be applied to determine whether to perform motion prediction.

It is necessary to determine whether to perform motion prediction between a block for which motion prediction is performed and a block for which motion prediction is not performed when having a predetermined feature vector value. For example, if performing motion prediction has better performance (smaller error energy) but is complex in computational complexity, a block with a predetermined feature vector may be applied by applying a predetermined reference value between the error energy and the computational complexity. It may be determined whether to perform this motion prediction.

In the case of a block having a specific vector, it may be determined whether to perform motion prediction using the motion vector.

If it is determined that the block does not perform the motion prediction, the probability value for not performing the motion prediction when the specific vector of the corresponding block is increased (step S250).

에서 움직임 예측을 수행하지 않는 블록이 될 경우의 조건부 확률이

라고 할 때 움직임 예측을 수행하지 않는 블록으로 판단된 경우 해당 확률 값을 높힐 수 있다.Specific vector value

Conditional probability for a block that does not perform motion prediction

If it is determined that the block does not perform the motion prediction can increase the corresponding probability value.

When it is determined that the block performs the motion prediction, the probability value of not performing the motion prediction when the specific vector value of the block is increased is increased (step S260).

에서 움직임 예측을 수행하는 블록이 될 경우의 조건부 확률이

라고 할 때 해당 확률 값을 높힐 수 있다.Specific vector value

Conditional probabilities in case of

If you can increase the probability value.

By learning whether to perform motion prediction when a specific feature vector value is adjusted by adjusting a probability value for a predetermined specific vector value, whether or not to perform motion prediction on a specific vector value can be determined by selecting a case having a larger probability value among conditional probability values.

Through a plurality of steps as described above, a conditional probability value for performing Bayesian classification may be calculated according to a specific vector.

The foregoing method is based on the assumption that the classifier uses a Bayesian classification method for convenience of description. However, when the classification method other than the above-mentioned Bayesian classification method is used in the motion vector estimation method using the classifier according to an embodiment of the present invention, the above-described method may be different, and the learning method for using such a different classification method is also provided. Included in the scope of rights.

3 is a flowchart schematically illustrating a motion vector estimation method using a classifier according to an embodiment of the present invention.

Referring to FIG. 3, a feature vector is calculated in a block (step S300).

A feature vector may be calculated to determine whether the block performs motion prediction. As described above, various values may be used for the feature vector.

The feature vector is classified (step S310).

The feature vector may have various values. For example, if the number of 1s included in a block after 1-bit conversion is a feature vector, the number of 1s may have many values depending on the number of pixels included in the block, and all of these values yield probability values. Since the computational complexity may increase, the value of the feature vector may be classified on a predetermined basis.

When the value of the feature vector does not vary, step S310 of classifying the feature vector may not be performed.

It is determined whether to perform motion prediction (step S320).

에서 움직임 예측을 수행하지 않는 블록이 될 경우의 조건부 확률

과 특정 벡터값

에서 움직임 예측을 수행하는 블록이 될 경우의 조건부 확률

을 비교하여

이

값보다 클 경우 움직임 예측을 수행하고

이

값보다 작거나 같을 경우 조기 종료를 수행할 수 있다. Specific vector value

Conditional Probability for Blocks that Do Not Perform Motion Prediction in

And specific vector values

Conditional Probability for Block Predicting Motion Prediction in.

By comparing

this

If it is greater than the value, the motion prediction is performed

this

If it is less than or equal to the value, early termination can be performed.

The motion prediction is performed (step S330).

As a result of the determination in step S320, when the probability when the motion prediction is performed on the feature vector is higher than the probability value when the motion prediction is not performed on the feature vector, the motion prediction may be performed.

The motion prediction is not performed (step S340).

As a result of the determination in S320, when the probability value when the motion prediction is performed on the feature vector is lower than or equal to the probability value when the motion prediction is not performed on the feature vector, the motion prediction may not be performed.

4 is a block diagram illustrating an apparatus for estimating a motion vector using a classifier according to an embodiment of the present invention.

Referring to FIG. 4, the motion vector estimator 400 using the classifier includes a 1-bit converter 410, a feature vector determiner 420, a preprocessor 430, a feature vector calculator 440, and a block classifier. 450 may be included.

The 1-bit converter 410 performs a 1-bit conversion using a filtering process using a filter kernel as shown in Equation 1 above, and a process of comparing the filtered result value with the original pixel value to a value of 1 or 0. can do. The 1-bit conversion methods disclosed in Equations 1 and 3 may be used as various 1-bit conversion methods as long as they are not out of the nature of the present invention as an example of 1-bit conversion.

The feature vector determiner 420 may select a feature vector to be used to classify the block.

The feature vector may be used by various variables representing the characteristics of the block, and one variable may be determined as the feature vector for classifying the block. When classifying blocks based on whether motion estimation is to be performed or not, an effective feature vector may vary according to characteristics of an image, and thus the feature vector determiner 420 may adaptively select a feature vector based on characteristics of an image. .

If the feature vector is used as a fixed variable, the feature vector determiner 420 may not be included in the motion vector estimator using the classifier.

The preprocessor 430 may perform preprocessing when necessary to clearly calculate the feature vector. For example, when performing binary block motion estimation, the number of 1s of pixel values of a block converted by 1-bit transform may be used as a feature vector. Therefore, the preprocessing step of calculating the number of 1s included in the binary block may be performed. If the preprocessing for calculating the feature vector is unnecessary, the preprocessor 430 may not be included in the motion vector estimating apparatus using the classifier.

In addition, the preprocessor 430 may classify the feature vectors into predetermined categories according to the feature vectors. For example, in H.264 / AVC, the quantization parameter has a value from 0 to 51. When the quantization parameter is a feature vector, the quantization parameter is classified into four predetermined intervals to reduce the computational complexity. Can be used to classify feature vectors. That is, when the quantization parameter values from 0 to 12 are included, the first feature vector 1 and the quantization parameter values from 13 to 25 may be classified as the second feature vector to reduce complexity when calculating probability values. If such classification is unnecessary in a specific vector, the preprocessor 430 may not be included in the motion vector estimating apparatus using the classifier.

The feature vector calculator 440 may calculate a feature vector for determining whether to predict motion in the block.

The feature vector may be, for example, a quantization coefficient, a screen size, a block size, a predicted motion vector size, a number of coded neighboring blocks, a number of skip coded neighboring blocks, and a binary block motion for a block to be encoded. Various values that can classify the characteristics of the block, such as the number of 1s obtained through the estimation and the difference between the pixel that passed through the band pass filter and the original pixel during binary conversion, may be used as the feature vector of the block.

The block classifier 450 may classify whether to perform motion prediction on the block based on the feature vector calculated from the feature vector calculator 440.

In the motion vector estimation apparatus using the classifier according to an embodiment of the present invention, a block may be classified into a block that does not perform motion prediction and a block that performs motion prediction.

As the classifier used in the block classifier 450, at least one of various classifiers such as a Bayesian classifier, a Perceptron classifier, and a support vector machine (SVM) may be used. It also belongs to the scope of the present invention.

The motion compensator 460 may generate a predictive block by estimating a motion vector through a binary motion prediction method based on a 1-bit transform method on a block classified as being performed by the block classifier 450. have.

In the case of a block classified as not performing the motion prediction by the block classifier 450, the motion compensator 460 estimates the motion vector of the current block from the neighboring block without directly estimating the motion vector, for example. can do.

Although the motion compensator 460 is illustrated as being located outside the motion vector estimating apparatus 400 for convenience of description, the motion compensator 460 and the motion vector estimating apparatus 400 may be implemented as one module. In other words, the present invention may be implemented as a plurality of modules or one module, as long as it does not depart from the essence of the present invention.

5 is a conceptual diagram illustrating an encoder including a motion vector estimating apparatus using a classifier according to an embodiment of the present invention.

FIG. 5 illustrates a configuration in which a classifier is included in a coder of a general H.264 / AVC for convenience of description, and according to an embodiment of the present invention in the configuration of an encoder other than H.264 / AVC, unless it is omitted from the essence of the present invention. Including a classifier according to the included in the scope of the present invention.

The DCT and quantization execution unit, the inverse DCT and inverse quantization execution unit, the deblocking filter, the motion compensation unit, and the like described in FIG. 5 are similar to the operation of each component of the general H.264 / AVC encoder.

According to the motion vector estimating apparatus using the classifier according to an embodiment of the present invention, the classifier 500 may be located at the front end of the motion compensator 510 or may be included in the motion compensator 510. As an example, the location of the classifier 500 is, for example, a flag provided at a previous stage of the deblocking filter 520 and having a flag indicating whether to perform motion prediction on a packet encoding the block. It can be expressed as information. That is, the position of the classifier may be variously located within the encoder, as long as it does not depart from the nature of the present invention.

The classifier 500 may determine whether to perform motion estimation on the corresponding block, and the block classified to not perform the motion estimation may not perform the motion prediction in the motion compensator 610.

6 is a table showing a result of performing a motion vector estimation method using a classifier according to an embodiment of the present invention.

MF-1BT represents a case of using the conventional 1-bit transform based motion prediction method based method, and 1BT-ET represents a motion prediction early termination method based on the existing 1 bit transform based motion prediction method.

Referring to FIG. 6, the BD-PSNR described in the horizontal axis of the table represents the difference of PSNR compared to the global search method based on the SAD method of each method, and Speed-up is a value for comparing the complexity of each method. The more you have, the lower the complexity.

As can be seen from the table, when the motion vector estimation method using the classifier according to the embodiment of the present invention is performed, the BD-PSNR value is similar to the PSNR value when the other conventional 1-bit transform based motion prediction method is performed. Thus, it can be seen that the image quality has a performance that does not deteriorate. However, the computational complexity is lower than that of the conventional 1-bit transform based motion prediction. That is, the computational complexity may be reduced even with similar reconstructed image quality.

Although described with reference to the embodiments above, those skilled in the art will understand that the present invention can be variously modified and changed without departing from the spirit and scope of the invention as set forth in the claims below. Could be.

Claims (15)

In the motion vector estimation method using a classifier,
Calculating a feature vector determining whether to predict motion in the block;
Based on a conditional probability value calculated as a result of learning using a plurality of images, the block is one of a first block performing motion prediction and a second block not performing motion prediction according to the conditional probability value of the calculated feature vector. Classifying; And
Generating a prediction block by performing motion prediction on the first block and generating a prediction block on the second block without performing motion prediction on the second block. The method of claim 1, wherein the motion vector estimation method using the classifier comprises:
And performing a one-bit transform on the reference frame and the current frame. The method of claim 2, wherein the generating of the prediction block by performing motion prediction on the first block and generating the prediction block without performing the motion prediction on the second block includes:
The motion vector estimation method using a classifier, characterized in that for performing the motion prediction based on a binary block motion estimation method for the reference frame and the current frame on which the 1-bit conversion is performed. The method of claim 1, wherein the motion vector estimation method using the classifier comprises:
Determining a feature vector to be used for classification; And
And performing a preprocessing to calculate the feature vector. delete delete The method of claim 1, wherein the feature vector,
Quantization parameter for the block to be encoded, the size of the frame, the size of the block, the size of the predictive motion vector, the number of encoded neighboring blocks in the screen, the number of skip encoded neighboring blocks, and the output through the 1-bit transform A motion vector estimation method using a classifier, characterized in that at least one of the number of 1, the difference between the pixel value passed through the band pass filter and the original pixel value during binary conversion. In the classifier for determining whether to estimate the motion vector of the block,
A feature vector calculator for calculating a feature vector for determining whether to calculate a motion prediction vector of the block; And
On the basis of the conditional probability value calculated as a result of learning using a plurality of images, whether to estimate the motion vector vector of the block including a block classification unit for classifying whether to perform the motion prediction in the block according to the conditional probability value for the feature vector Classifier to determine. The method of claim 8, wherein the classifier for determining whether to estimate the motion vector of the block,
A classifier for determining whether to estimate the motion vector of the block further comprising a one-bit converter for performing one-bit conversion of the current frame and the reference frame by performing a predetermined filtering. The method of claim 8, wherein the classifier for determining whether to estimate the motion vector of the block,
A feature vector determiner that determines a feature vector to be used to classify the block; And
A classifier for determining whether to estimate the motion vector of the block further comprising a preprocessor for performing the preprocessing for calculating the feature vector. delete delete delete delete delete

Images