Video compression system

This nonprovisional application claims domestic priority to prior provisional application Ser. No. 60/312,941, filed Aug. 16, 2001, and to provisional application Ser. No. 60/316,316, filed Aug. 31, 2001. Both provisional applications are incorporated herein by reference.

U.S. Pat. Nos. 6,178,205; 6,167,164; and 6,188,799 entitled “AN EFFICIENT REAL TIME ALGORITHM TO SIMULTANEOUSLY REDUCE BLOCKING AND RINGING ARTIFACTS OF COMPRESSED VIDEO” by Min-Cheol Hong, Chang Mo Yon and Young Man Park, assigned to Digital Media Research Lab IEEE 1999.

The following computer software files, submitted herewith on one (1) CD-R compact disc, are incorporated by reference herein:

This application specifically relates to a method and apparatus for reducing the decode complexity of two dimensional inverse transforms on a vector process.

A typical digital video decoding system involves the following steps (among others).

For each block in a frame:

A) Extract quantized transform coefficients from the compressed bit-stream B) Perform inverse quantization to reconstruct the transform coefficients C) Perform an inverse transform (typically an IDCT) on the coefficients D) Add the resultant values to a block predictor E) Output the block results

The 2-dimensional inverse transform functions typically take a large portion of the time to decode a frame due to their complexity.

The invention described here attempts to reduce the decoder complexity on vector processing machines that are capable of doing the same operation to multiple values stored sequentially in a machine's registers by lowering the complexity of the 2 dimensional transform.

A 2-dimensional separable inverse transform performed on a block typically involves performing the following steps:

a) For each row of the block:

Perform the same 1-dimensional inverse transform on the transform coefficients.

b) For each column of the block column (resulting from (a)):

Perform the same 1-dimensional inverse transform on the transform coefficients.

Since the 1-dimensional inverse transform usually involves performing exactly the same operations on a number of rows or columns in the block, vector processors are often used to reduce the decoding time. This is typically accomplished by filing vector processing registers with a value from each of N rows in the block (see diagram). The operations of the inverse transform are then performed on the N rows in parallel. And then the vector processing registers are filled with values from each of the N columns in the block and the inverse transform is then performed on the N columns in parallel.

In order to fill the vector processing registers quickly with different values from each row a programmer typically has two options:

a) Transpose the coefficients so that the transform coefficients appear in the order that matches the vector processor and load them directly into the registers. b) Fill the vector registers one value at a time with the coefficient data.

Choice (a) requires numerous operations to perform the transpose and choice (b) requires numerous bit-mask AND/OR operations to place each coefficient into the register.

This invention attempts to address these issues. To do so:

a) Fill an entire frame's coefficient buffers with 0 before the start of coefficient decoding. b) Extract the transform coefficient for a token from the bit-stream. c) If the coefficient value is non-zero place it in the transposed order that best suits the target vector processor.

This works better than performing the transpose or filling the vector registers one at a time as part of the inverse transform because there are typically many more zero than non-zero coefficients. As such we make only a few positional changes, and avoid doing one full transpose altogether.

Additional savings are achieved by an embodiment of this technique whereby the same coefficient for every block in the image is encoded in the bit-stream before moving on to the next coefficient. In this way, one look up can be performed for each coefficient in the transform to determine where to place the transform coefficient.

Since each processor has vector registers with different cardinality it is necessary to place the coefficients in the order that best suits the processor being used.

The specific embodiment uses an IDCT transform but the technique is equally applicable to any separable 2-dimensional transform, for example, the discrete wavelet transform or the generalized orthogonal transform.

FIG. 1 is a block diagram of the embodiment.

FIG. 2 is an overview block diagram of the embodiment.

The invention includes enhanced video processing and compression and is further described hereinafter.

The encoder uses a motion estimator, block based 8×8 Discrete Cosine Transform (DCT), a quantizer, a variable length encoder, and a loop filter for smoothing block edges in the reconstruction buffer. The decoder uses a variable length decoder and inverse quantizer, a motion compensator and a loop filter for smoothing block edges.

Two separate image artifacts are produced as a result of the quantization step. A blocking artifact is produced when quantization of the DCT coefficients in adjacent blocks produces pixel values on the shared block edge that differ on either side of the edge by a greater amount than in the original image. A ringing or mosquito artifact results from the quantization of higher frequency components of the transform around strong edges in the image. This means that the transform basis vectors do not reinforce and cancel correctly, producing edges in the reconstruction near to the strong edge that were not present in the original image.

The current invention embodies two separate but dependent filters that attempt to remove these image artifacts in a manner that is low on decoder complexity:

a) a de-blocking filter that reduces the blocking artifacts described above b) an edge-enhancement and de-ringing filter that attempts to reduce the mosquito noise or ringing artifacts at the same time it sharpens real text images.

A block diagram of the vectorized de-blocker is shown in FIG. 2 .

The de-blocking filter of FIG. 2 works as follows:

Define an array, BlockSadArray, that contains one entry for each 8×8 block in the image. Initialize each entry of BlockSadArray to the value 0. For every border between two 8×8 blocks of the image:

At each of the eight positions along the border, examine the 10 pixels that lie in a direction perpendicular to the border, five from each block on either side (see diagram). Label these 10 pixels x0 . . . x4 and x5 . . . x9, respectively. Computer the following sums at each of the eight positions on the border:

If the following condition is met:

Side 1 Sad< 3*T1/2 and

Side 2 Sad< 3*T2/2 and

Abs ( x 4 −x 5)< T 3,

then replace values x1 to x8 with low pass filtered values x1 to x8 as follows:

x 1′=( x 0+ x 0+ x 0+ x 1*2+ x 2+ x 3+ x 4+4)/8

x 2′=( x 0+ x 0+ x 1+ x 2*2+ x 3+ x 4+ x 5+4)/8

x 3′=( x 0+ x 1+ x 2+ x 3*2+ x 4+ x 5+ x 6+4)/8

x 4′=( x 1+ x 2+ x 3+ x 4*2+ x 5+ x 6+ x 7+4)/8

x 5′=( x 2+ x 3+ x 4+ x 5*2+ x 6+ x 7+ x 8+4)/8

x 6′=( x 3+ x 4+ x 5+ x 6*2+ x 7+ x 8+ x 9+4)/8

x 7′=( x 4+ x 5+ x 6+ x 7*2+ x 8+ x 9+ x 9+4)/8

x 8′=( x 5+ x 6+ x 7+ x 8*2+ x 9+ x 9+ x 9+4)/8

The invention encompasses the following novelty:

1: The decision as to whether or not to perform the de-blocking operation can easily be computed for several positions at a time using a vector processor (such as Intel Corporation Inc.'s MMX architecture) with very few operations. 2: The filter applied itself is much simpler than most and requires far fewer operations than those found in the prior art.

b) The edge enhancement and de-ringing filter works as follows:

For each 8×8 block in the image:

Let BSAV be the BlockSadArray value corresponding to the current block. Define three threshold values Medium Thresh, High Thresh, Highest Thresh that are computed as a function of the quantization level such that Highest Thresh>High Thresh>Medium Thresh IF (BSAV>Highest Thresh)

Apply the strong de-ringing filter in place to the block (described below) multiple times (based on how high BSAV is)

ELSE IF (BSAV>High Thresh)

Apply the strong de-ringing filter to the block (described below)

ELSE IF (BSAV>Medium Thresh)

Apply the weak de-ringing filter to the block (described below).

The strong de-ringing filter works as follows:

Calculate a maximum blurring modifier (HighModifier) and maximum sharpening modifier (LowModifier) by looking up a value based upon the level of quantization applied to the coefficients.

For the j th pixel in the i th image row, P i,j , do the following:

Calculate the following four values:

M 0 =DeringModifier(P i,j , P i−1,j )

M 1 =DeringModifier(P i,j , P i+1,j )

M 2 =DeringModifier(P i,j , P i,1−j )

M 3 =DeringModifier(P i,j , P i,−j+1 )

Where DeringModifier(X,&) is defined as follows:

Value=Constant+QuantizationLevel−Abs(X−Y) IF (Value<LowModifier)

Value=LowModifier

IF (Value>HighModifier)

Value=HighModifier

Replace Pixel P i,j with P i,j ′ computed as follows:

P i,j ′=( M 0 *P i−1j +M 1 *P i+1j +M 2 *P ij−1 +M 3 *P i,j+1 +(128−( M 0 +M 1 +M 2 +M 3 )* P i,j )/128

The deblocker uses simple linear calculations and one dimensional filters to remove deblocking artifacts and to gather information used by the deringer filter to determine how many iterations to apply its filter. The deringing filter encompasses a pixel to pixel spatially adaptive filter that can both blur and sharpen. It does so by collecting the difference between neighboring pixels into a two dimensional array, applying a simple function to this array and then using the result as a convolution kernel. Since the function can produce both positive and negative tapes, the filter can perform both deringing and sharpening.

The Weak de-ringing filter works exactly the same as above except that the High and Low Modifiers are smaller magnitude numbers and the DeringModifier Function works as follows:

Value=Constant+QuantizationLevel−2*Abs(X−Y) IF (Value<LowModifier)

Value=LowModifier

IF (Value>HighModifier)

Value=HighModifier

FIG. 1 describes the simplified process by which the 2d orthogonal transform works. The basic process is as follows: First initialize all of the coefficients for every block in a group of blocks to 0. Then determine the position of the best place in memory to place the first coefficient given the vector processor. Next extract the first coefficient for each block in a group of blocks, and write any non 0 coefficient to the appropriate location. Then determine the position of the best place in memory to place the second coefficient given the vector processor. Next extract the second coefficient for each block in a group of blocks, and write any non 0 coefficient to the appropriate location. Repeat this same process for each of the subsequent coefficients.

Enclosed with this application is a CD-Rom with the preferred embodiment illustrated by providing a listing in Source Code of the method, system and steps of this invention. Following a reading of said CD-Rom, the invention herein is again summarized.

The novelty of this approach in comparison to others found in prior art includes:

1: The selection of different kernel modifiers based upon the difference between the target pixel and its neighbors which is an extremely simple and solely integer based linear calculation that may be efficiently implemented on a vector processor. 2: The ability of the filter to both sharpen real image edges and blur away ringing artifacts and mosquito noise. 3: The usages of completely different de-ringing functions based upon sum absolute difference value information gathered by the de-blocker. 4: The ability to run the de-ringing functions iteratively based on the BlockSad value for the block.

In summary, the following table identifies some of the advantages, features and benefits of this invention.

Having described this invention with regard to specific embodiments, it is to be understood that the description is not meant as a limitation since further modifications and variations may be apparent or may suggest themselves to those skilled in the art. It is intended that the present application cover all such modifications and variations as fall within the scope of the appended claims.