Learning Filters For Enhancing The Quality Of Block Coded Still And Video Images

• US 20090034622A1
• Filed: 08/01/2008
• Published: 02/05/2009
• Est. Priority Date: 08/01/2007
• Status: Active Grant

• Current Assignee: Her Majesty The Queen In Right Of Canada, As Represented By The Minister Of Industry, Through The Communications Research Centre Canada
• Original Assignee: Her Majesty The Queen In Right of Canada As Represented By The Minister of National Defence, Through The Communications Research Centre Canada
• Inventors: Huchet, Gregory, Wang, Demin, Vincent, Andre, Chouinard, Jean-Yves
• Granted Patent: US 8,204,128 B2
• Time in Patent Office: Days
• US Class Current: 375/240.160
• CPC Class Codes: H04N 19/117 Filters, e.g. for pre-processing or post-processing sub-band filter banks H04N19/635
• CPC Class Codes: H04N 19/154 Measured or subjectively estimated visual quality after decoding, e.g. measurement of distortion use of rate-distortion criteria H04N19/147
• CPC Class Codes: H04N 19/176 the region being a block, e.g. a macroblock
• CPC Class Codes: H04N 19/182 the unit being a pixel
• CPC Class Codes: H04N 19/46 Embedding additional information in the video signal during the compression process H04N19/517, H04N19/68, H04N19/70 take precedence
• CPC Class Codes: H04N 19/61 in combination with predictive coding
• CPC Class Codes: H04N 19/82 involving filtering within a prediction loop
• CPC Class Codes: H04N 19/86 involving reduction of coding artifacts, e.g. of blockiness

Abstract

The invention relates to a method and an apparatus for reducing blocking artifacts in block-wise coding of still and video images. A learning filter generator is provided at the image encoder for generating a set of filters and associated filtering rules for filtering cross-boundary image patterns based on representative original and decoded training images using a supervised machine learning algorithm. An adaptive filter at the image decoder receives the generated filters and associated filtering rules and performs locally adaptive filtering in accordance with the received filtering rules.

Claims

1. A method for block-wise encoding still or video images comprising the steps of:

a) encoding input image data composed of pixels having original pixel values using a block-wise encoder to obtain encoded image data;

b) decoding at least a subset of the encoded image data to obtain a decoded image composed of pixels having decoded image values;

c) identifying in the decoded image a plurality of coding block boundaries;

d) forming training data comprising a plurality of cross-boundary filter support blocks (CBFSBs), each CBFSB comprising a group of N>

1 adjacent pixels of the decoded image including adjacent pixels on both sides of a block boundary;

e) generating a set of classification rules based on the training data for classifying each of the CBFSBs into at least one of a plurality of classes, so that each class is adaptively defined by a respective set of the classification rules;

f) for each class defined in step (e), generating a set of filter coefficients associated with the class and defining a class-optimized filter that provides a prediction value of a selected CBFSB pixel in dependence upon decoded values of CBFSB pixels based on an original value of the selected CBFSB pixel for each CBFSB in said class; and

,

g) providing the sets of filter coefficients for each class and the classification rules as an output for subsequent adaptive class-based filtering of the decoded image data for suppressing blocking artefacts.

2. A method according to claim 1, wherein step (e) is performed using a supervised machine learning algorithm.

3. A method according to claim 2, wherein the supervised machine learning algorithm is one of:

an M5 algorithm, an M5′

algorithm, a multilayer perception algorithm, or a linear regression algorithm.

4. A method according to claim 2, wherein the supervised machine learning algorithm comprises iterative partitioning of the training data.

5. A method according to claim 1, further comprising the steps of:

h) receiving the encoded image data, the filter coefficients for each class, and the classification rules;

i) decoding the encoded image data to obtain received image data;

j) selecting a target pixel in the received image data;

k) identifying in the received image data a first nearest block boundary in the vicinity of the target pixel;

l) selecting a first CBFSB of size N that includes the target pixel and crosses said first nearest block boundary;

m) classifying the first CBFSB into one of the classes defined in step (e) using the received classification rules, and selecting a first set of filter coefficients associated therewith from the received sets of filter coefficients;

n) applying the first set of filter coefficients to the first CBFSB to obtain a first filtered pixel value for the target pixel.

6. A method according to claim 5, further comprising the steps of:

k2) identifying in the received image data a second nearest block boundary in the vicinity of the target pixel;

l2) selecting a second CBFSB of size N that includes the target pixel and crosses said second nearest block boundary;

m2) classifying the second CBFSB into one of the classes defined in step (e) using the received classification rules, and selecting a second set of filter coefficients associated therewith from the received sets of class-optimized filter coefficients;

n2) applying a corresponding set of class-optimized filter coefficients to the second CBFSB to obtain a second filtered pixel value for the target pixel; and

,

o) averaging the first and second filtered pixel values to obtain a new filtered value of the target pixel.

7. A method according to claim 6, wherein the first and second nearest block boundaries are vertical and horizontal boundaries of a same coding block.

8. A method according to claim 1, wherein the input image data comprises a sequence of video images, the method comprising the steps of:

selecting the subset of the encoded image data that corresponds to a representative image from the sequence of video images, and

applying the classification rules and the class-optimized filters for filtering other decoded images of the input sequence of video images.

9. A method according to claim 1 for use in a video encoder utilizing motion estimation and motion compensation, wherein the input image data comprises a sequence of input video images, wherein step (b) comprises generating a sequence of decoded images corresponding to the sequence of input video images, the method further comprising the step of

applying the classification rules and the class-optimized filters for filtering each decoded image from the sequence of decoded images prior to performing the motion compensation thereon.

10. A method according to claim 1 that is used in a video encoder utilizing motion estimation and motion compensation, wherein the input image data comprises a sequence of input video images, wherein step (b) comprises generating a sequence of decoded images corresponding to the sequence of input video images, the method further comprising the step of

applying the classification rules and the class-optimized filters for filtering each decoded image from the sequence of decoded images after performing the motion compensation thereon.

11. A method according to claim 1, wherein steps (e) and (f) comprise the steps of:

e1) selecting a pixel position i within each CBFSB for filtering;

e2) identifying from the input image data original pixel values p^o_i corresponding to the i-th pixel position in each CBFSB;

e3) forming a plurality of training instances I, each training instance I comprising N decoded pixel values of one of the CBFSB and the original pixel value p^o_i for the i-th pixel position in said CBFSB;

e4) processing the plurality of training instances I to obtain classification rules for classifying the training instances I according to a selected classification criterion using a supervised machine learning algorithm;

f1) for each of the classes identified in step e4), determining the respective class-optimized filter providing the predicted values of pixels at the i-th CBFSB location; and

,

f2) repeating steps (e1)-(e4) for at least one other pixel position j within each CBFSB.

12. A method according to claim 11, wherein the selected classification criterion comprises maximizing a standard deviation reduction value for a group of CBFSBs.

13. A method according to claim 12, wherein step (e) includes splitting a group of CBFSBs in two or more sub-groups in dependence on a pixel value of one of the CBFSB pixels.

14. A method according to claim 1, wherein the classification rules comprise ranges of pixel values associated with pixel positions in an CBFSB.

15. A method according to claim 1, wherein step (f) is performed using a linear regression algorithm, and comprises computing a difference between the prediction value and the original pixel value to determine a prediction error for each CBIS of a class, and selecting a set of filter coefficients that minimizes or maximizes a merit function based on the prediction errors for each of the CBIS of the class.

16. A method for block-wise decoding still or video images encoded by the method of claim 1, comprising the steps of:

receiving the encoded image data, the set of classification rules for classifying cross-boundary filter support blocks (CBFSB) into at least one of the plurality of classes, and the sets of filter coefficients associated with said classes;

decoding the encoded image data with a block-wise decoder to obtain the received image data;

identifying a plurality of coding block boundaries in the received image data;

for a target pixel in the received image data, performing the steps of;

identifying a CBFSB in the received image data that comprises said target pixel and crosses one of the coding block boundaries;

classifying said CBFSB into one of the adaptively defined classes using the received set of classification rules; and

, applying to the identified CBFSB in the received image data the filter coefficients associated with said class to obtain a filtered value of the target pixel.

18. An image encoder of claim 16 comprising a coding block detector for identifying a plurality of coding block boundaries in the decoded image.

19. An image decoder for decoding still or video images that have been encoded with the image encoder of claim 16, comprising:

a data port for receiving the encoded image data, the classification rules for classifying cross-boundary filter support blocks (CBFSB) into at least one of the classes adaptively defined at the image encoder of claim 16, and the sets of filter coefficients associated with said classes;

a block-wise decoder for decoding the encoded image data to obtain the received image data;

an adaptive filter coupled to the block-wise decoder for class-based filtering of target pixels of the received image data, comprising;

an CBFSB sorting section configured to identify, for each of the target pixels in the received image data, an CBFSB that comprises said target pixel and crosses a coding block boundary, and to associate said CBFSB with one of the adaptively defined classes using the received set of classification rules; and

,

a filtering section for applying to each of the CBFSBs identified in the received image data one of the sets of class-optimized filter coefficients according to a respective class associated therewith.

20. An image decoder of claim 18 comprising a block detector for identifying coding block boundaries in the received image data.

17. An image encoder for encoding still or video images, comprising:

a block-wise encoder for generating encoded image data from input image data comprised of pixels having original pixel values using coding blocks having coding block boundaries;

a block-wise decoder operatively connected to the block-wise encoder for receiving at least a subset of the encoded image data and for generating therefrom a decoded image comprised of pixels having decoded pixel values;

a training data generator for forming a training dataset comprising a plurality of cross-boundary filter support blocks (CBFSB), each CBFSB comprising a group of N>

1 adjacent pixels of the decoded image including adjacent pixels on both sides of a block boundary;

a learning classifier operatively connected to the input port and the training data generator for receiving the input image and the corresponding decoded image for generating therefrom

a set of classification rules for classifying each of the CBFSBs into at least one of a plurality of classes that are thereby adaptively defined, and

at least one set of class-optimized filter coefficients for each adaptively defined class so that filtered values of CBFSB pixels obtained using said filter coefficients approximate original pixel values of respective pixels for all CBFSB associated with said class; and

,

a data port coupled to the block encoder and the learning classifier for providing the encoded image data, the classification rules and the set at least one set of the class-optimized filter coefficients as an output for subsequent decoding and class-based pixel filtering for reducing blocking artefacts.