Universal MPEG decoder with scalable picture size

<b>CROSS-REFERENCE TO RELATED APPLICATIONS</b>

0001The present application claims priority from Provisional Application Serial No. 60/009,161 entitled Universal MPEG Decoder with Scalable Picture Size filed December 22, 1995.

<b>FIELD OF THE INVENTION</b>

0002The present invention relates to a Discrete Cosine Transform-based decoding apparatus for decoding video data streams in multimedia computer systems. More specifically, this invention relates to an MPEG video stream decoder for decoding video data encoded in any one of three color space formats. Motion video is encoded according to specific standards and specific color space formats. A universal decoder may simplify decoding and output of a video data stream in different formats.

<b>BACKGROUND OF THE INVENTION</b>

0003Computers with multimedia capabilities have become commonplace. Multimedia computers are loosely defined as computers with display capability for text, high resolution graphics, high resolution images, high dynamic range audio and high resolution motion video. One problem accompanying multimedia computing may be large image frame sizes. Large image frames stored in primary or secondary memory may be retrieved from memory, transferred across system busses and displayed via display controllers. With larger images, more data may be transferred and it may take additional time for large images to be displayed. Problems may be magnified for display of multiple successive image frames associated with motion video. Fast video frame updates must occur frequently enough to provide video free from anomalies such as delay, slowed motion, or choppiness associated with slow or erratic updating.

0004The standard in high resolution still frame image display may be an image with 1024 pixels by 768 lines of display resolution with 24 bits of color information per pixel. With no compression or other encoding, a standard frame approaches a size of 20 megabytes. With full motion digital video requiring a minimum frame rate of about 30 frames per second to simulate live motion, full-size, full-resolution, non-encoded motion video may be impractical on existing computers with limited bus transfer capability. In order to successfully display high resolution motion video on multimedia computers it may be necessary to employ a scheme to compress and encode motion video data to minimize frame size for intrabus transfer. Image frames may be fully decompressed and displayed at full resolution in a display controller after bus transfer. It may also be necessary to set a maximum standard window or frame size for efficient display of motion video. The combination of optimum frame size and compression encoding method may make display of motion video on multimedia computers possible.

0005MPEG has become a motion video compression standard widely accepted in both hardware and software multimedia communities as a compression and decompression standard of preference for motion video frames. MPEG derives its name from International Standards Organization's (ISO) Motion Picture Encoding Group (MPEG). MPEG specifications call for compression of consecutive motion video frames accomplished by making the first frame in a video sequence and periodic frames thereafter, reference or base frames. MPEG specifications then call for computing a mathematical difference between a present frame and a previous base frame, storing only a computed difference between the two. Motion compensation may also be performed since an object in one frame which would normally register no differences in subsequent frames when still may register a difference attributable only to motion.

0006MPEG computes differences between base frames and subsequent inter-frames by breaking frames into 8- by 8-pixel blocks and matching pixels from each block to corresponding pixel locations in the base frame. Differences are compressed using a method called DCT (Discrete Cosine Transform). Once computed, DCT coefficients may be Huffman-coded to produce a final block size which may often be one-tenth to one-twentieth of original block size. The DCT may be integral to the preferred embodiment of the present invention. A detailed description of the DCT is presented here as background.

0007If a DCT block size is assumed to be N in one direction, then a type-2 DCT (X_IIE) used for computing a forward DCT for a block size of NxN may be given by:<maths id="math0001" num="(1)"><math display="block"><mrow><msub><mrow><mtext>X</mtext></mrow><mrow><mtext>IIE</mtext></mrow></msub><msup><mrow><mtext>{X(n)} = (2/N)</mtext></mrow><mrow><mtext>1/2</mtext></mrow></msup><mtext>k(m)</mtext><apply><sum /><lowlimit><mtext>n = 0</mtext></lowlimit><uplimit><mtext>N-1</mtext></uplimit><mrow><mtext>x(n)cos(Πm(n+1/2)/N)</mtext></mrow></apply><mtext>   m = 0,1...N-1</mtext></mrow></math><img file="EP0781052A2_D0001.tif" /></maths> where x(n) is a pixel in block N x N.

0008A weighing function may be defined as follows:<maths id="math0002" num=""><img file="EP0781052A2_D0002.tif" /></maths>

0009To understand Remez Exchange low-pass Finite Impulse Response (FIR) filtering which may be used in the preferred embodiment of the present invention, a discussion of general characteristics of FIR filters may be helpful.

0010An FIR filter may sometimes be called a moving-average filter. An FIR filter may also be called a non-recursive filter or a convolution filter. The duration or sequence length of the impulse response of these filters is by definition finite; therefore, an output function may be written as a finite convolution sum as follows:<maths id="math0003" num="(2)"><math display="block"><mrow><mtext>y(n) = </mtext><apply><sum /><lowlimit><mtext>m = 0</mtext></lowlimit><uplimit><mtext>N-1</mtext></uplimit><mrow><mtext>h(m) x (n-m)</mtext></mrow></apply></mrow></math><img file="EP0781052A2_D0003.tif" /></maths> where x(n) are input vectors and h(m) is the impulse response.

0011An FIR filter may be interpreted as an extension of a moving sum or as a weighted moving average. In any sequence of random numbers, such as prices from daily stock market prices for a particular stock, it may be desirable to remove erratic variations to discover longer trends. Each number may be replaced by a value corresponding to an average of itself and the preceding three values. Variations within a four day period may be "averaged" out and "longer" term variations remain.

0012For illustration, we consider a signal x(n), which may contain a sum of a linear term K₁n, and an undesired oscillating signal, such that<maths id="math0004" num="(3)"><math display="block"><mrow><msub><mrow><mtext>x(n) = K</mtext></mrow><mrow><mtext>1</mtext></mrow></msub><msub><mrow><mtext>n + K</mtext></mrow><mrow><mtext>2</mtext></mrow></msub><mtext> cos(Πn)</mtext></mrow></math><img file="EP0781052A2_D0004.tif" /></maths>

0013If we assume two elements in a FIR filter, then h(n) may be<maths id="math0005" num=""><img file="EP0781052A2_D0005.tif" /></maths>

0014After two iterations, the output is a linear term with a delay of one-half samples and no oscillation illustrating the moving average principal of an FIR filter.

0015Upscaling or interpolation of a motion video frame containing image data involves introducing zeros as data values for even samples. After upscaling image data, a low-pass filter stage may be applied to image data signal to average data evenly over a resized block which has undergone interpolation and decimation. The filter may be an even-length symmetric filter.

0016In low-pass filtering based on a Remez Exchange algorithm, five parameters may be of interest in a symmetrical filter: <ul id="ul0001" list-style="none" compact="compact"><li>1) N - Filter length</li><li>2) f_p - The edge of the pass band specified as a fraction of sampling frequency.</li><li>3) f_s - The edge of the stop band specified as a fraction of sampling frequency.</li><li>4) δ₁ - The deviation from unity in the pass band.</li><li>5) δ₂ - The deviation from zero in the pass band.</li></ul>

0017A low-pass Remez filter may be an even length symmetric filter whose tap size is 32. A tap size of 32 may accommodate DCT block-sizes which may be either 16x8 or 16x16. More taps in the Remez Exchange filter may result in better convolution of filter coefficients and more accurate reproduction of upsampled and downsampled pixels.

0018A Remez low-pass filter may be generally defined as h(n), <maths id="math0006" num=""><math display="inline"><mrow><mtext>N=-L/2.. 0,....L/2-1</mtext></mrow></math><img file="EP0781052A2_D0006.tif" /></maths>. The Remez filter right-half is defined by<maths id="math0007" num=""><img file="EP0781052A2_D0007.tif" /></maths> where L may be the number of filter coefficients for h(n) and N may be the block size of the IDCT.

0019MPEG implementations require several important functions difficult to implement on a general-purpose CPU. Motion estimation and compensation require block comparisons between corresponding base frame block locations and present frame block locations. 8 by 8 pixel DCT block location comparisons between intraframes and base frames may not correspond one to one due to motion. The DCT specification in MPEG, like other compression methods, may be more effective with highly correlated data. Thus, even though motion may have occurred in intraframes relative to base frames, computationally intensive search procedures incorporating motion detection may be used to find such redundancies.

0020Existing hardware MPEG decoders (e.g., CL450) decode a compressed bitstream into a single display format such as YUV 4:2:0, YUV 4:2:2, and YUV 4:4:4 format depending on a particular compression format used upon encoding. Encoding formats may be geared toward anticipated display device capabilities or may be chosen to provide a format appropriate for compatibility with the broadest base of display devices. In practice, display devices may most likely be VGA terminals. Most VGA terminals may display video in YUV 4:2:2 color space format. Other potential destination devices may include a D-1 recorder typically employing YUV 4:4:4 format.

0021YUV color space is the basic color space used by PAL (Phase Alternation Line), NTSC (National Television System Committee) and SECAM (Sequential Couleur Avec Memoire) color video standards. The Y component (luminance) stands for intensity information while U and V components (chrominance) stand for color information. MPEG compression and decompression standards specify color space encoding in YUV color space. There are three primary YUV formats: YUV 4:2:0, YUV 4:2:2, and YUV 4:4:4. Such color space formats may describe the proportion of Y,U, and V component information in each byte or collection of bytes of display information.

0022Motion video may comprise encoded data corresponding to each frame of video data in a motion video sequence. Each frame may comprise three rectangular matrices representing luminance (Y) and two chrominance (CbCr) values. The Y matrix may have an even number of rows and columns. In MPEG-1, the Y matrix may describes the intensity component of a frame of size 352 pixels by 240 lines. The chrominance components may be one of several configurations depending on the encoding format. Figure 1 illustrates three primary color space encoding formats used in MPEG compression.

0023In YUV 4:2:0 format illustrated in Figure 1a, chrominance component matrices 111, and 112 are one half in size of Y matrix 110 in horizontal and vertical directions. In YUV 4:2:2 format illustrated in Figure 1b, chrominance component matrices 121, and 122 are one half in size of Y matrix 120 in the horizontal direction and the same size as Y matrix 120 in the vertical direction. Finally, in YUV 4:4:4 format illustrated in Figure 1c, chrominance matrices 130, and 131 are specified as being the same size as Y matrix 130 in both horizontal and vertical directions.

<b>SUMMARY OF THE INVENTION</b>

0024A universal MPEG decoder decodes a video bitstream representing video frame blocks MPEG encoded in one of three primary color space formats. The decoder of the present invention may be selectively programmed to support YUV 4:2:0, YUV 4:2:2, and YUV 4:4:4 input and output display formats. Any input format may be decoded and output in any output format. Moreover, the decoder of the present invention performs dynamic image frame-size scaling.

<b>BRIEF DESCRIPTIONS OF THE DRAWINGS</b>

0025Figure 1a is a diagram illustrating YUV 4:2:0 color space format.

0026Figure 1b is a diagram illustrating YUV 4:2:2 color space format.

0027Figure 1c is a diagram illustrating YUV 4:4:4 color space format.

0028Figure 2 is a block diagram illustrating elements of a simple MPEG Decoder.

0029Figure 3 is a block diagram illustrating elements of the IDCT block of the present invention.

0030Figure 4 is a diagram illustrating the pass band and stop band of a Remez Exchange filter superimposed upon a unit circle.

0031Figure 5 is a graph illustrating the frequency response of a Remez Exchange filter.

0032Figure 6 is a diagram illustrating an image processed using prior art spatial domain interpolation.

0033Figure 7 is a diagram illustrating an image processed using DCT domain interpolation of the present invention.

0034Figure 8 is a graph illustrating a signal filtered using the present invention.

DETAILED DESCRIPTION OF THE INVENTION

0035The descriptions herein are by way of example only illustrating the preferred embodiment of the present invention. However, the method and apparatus of the present invention may be applied in a similar manner in other embodiments without departing from the spirit of the invention.

0036Figures 1a, 1b, and 1c are diagrams illustrating three different formats describing color space in a motion video frame. A motion video frame may consist of three rectangular matrices representing luminance Y and two chrominance (CbCr) or U and V values. Y matrices 110, 120, and 130 may have an even number of rows and columns. Chrominance component matrices 111 and 112 may be one half in size of Y matrix 110 in horizontal and vertical directions in YUV 4:2:0 color space format. Figure 1a is a diagram illustrating YUV 4:2:0 color space format. In YUV 4:2:2 format, chrominance component matrices 121, and 122 may be one half in size of Y matrix 120 in the horizontal direction and of the same size in the vertical direction. Figure 1b is a diagram illustrating YUV 4:2:2 color space format. In color space format YUV 4:4:4, chrominance matrices 131, and 132 may be the same size as Y matrix 130 both in the horizontal and vertical directions. Figure 1c is a diagram illustrating YUV 4:4:4 color space format.

0037Figure 2 is a block diagram illustrating elements of an MPEG Decoder. Coded data 201 is input to Variable Length Decoding block 202. Data decoded in Variable Length Decoding block 202 is output to Inverse Scan block 203. Data processed by Inverse Scan block 203 is output to Inverse Quantizer block 204. Inverse Quantized data processed in Inverse Quantizer block 204 is output to Inverse Discrete Cosine Transform (IDCT) block 205. Data processed by IDCT block 205 is output to Motion Compensation block 206.

0038Motion estimation may be performed in Motion Compensation block 206 on luminance and chrominance components in an MPEG encoder. Motion vectors may be computed for luminance components only. Image scaling only may affect motion compensation for chrominance components. When motion compensation is decoded in an MPEG decoder, motion vectors for chrominance components may be derived from luminance components and a scaling factor. For example, an image upscaled from 4:2:0 to 4:4:4 may have the same motion vectors for chrominance components Cb and Cr as for Y. Data processed by Motion Compensation block 206 is output as Decoded Data Out 208. Video frames occurring as Decoded Data Out 208 may be stored in Frame Store Memory 207 and may be used by Motion Compensation block 206 to compare the present frame with previous frames. Because image scaling may require minimal color space formats to be upscaled to comprehensive color space formats, Frame Store Memory 207 may have to be large enough to accommodate image frames corresponding to YUV 4:4:4 format.

0039Figure 3 is a block diagram illustrating elements of IDCT block 205 of Figure 2. Inverse Quantized DCT coefficients 301 are input from Inverse Quantizer block 204 to Anti-aliasing Anti-imaging block 303 and Up-sampling/Down-sampling block 304. Coefficients are output to IDCT block 305. Programmable Coefficients 306 may be input to IDCT block 305 prior to output of a video stream to Motion Compensation block 307. Desired Picture Format 302 may be input to IDCT block 205 to specify the color format of the output of the decoder of the present invention.

0040Figure 4 is a diagram illustrating pass band and stop band of the Remez Exchange filter superimposed upon a unit circle of radius 1 and circumference of 2Π. The pass-band may be between 0 and 1/2Π and between 3/2Π in the 2Π unit circle. Pass band PB 401 and stop band SB 402 of a Remez Exchange filter are illustrated in the response plot of Figure 4. Because the filter is symmetric, PB 401 includes all the signals with frequency's between 0 and 1/2Π and 0 and -1/2Π(3/2Π). Similarly, SB 402 is between 0.5Π and -0.5Π. Gain of the filter is unity in PB 401 and attenuation in SB 402 is -30db.

0041The following are computed Remez Exchange filter coefficients.    4.4996e-01 2.3573e-02    1.4953e-01 2.0967e-02    -8.9156e-02 -1.8786e-02    -6.3129e-02 -1.4276e-02    4.8479e-02 9.9493e-03    3.9131e-02 2.2767e-02    -3.2452e-02 -3.5774e-02    -2.7510e-02 -1.7824e-02

0042Filter coefficients shown above may represent h(n) for <maths id="math0008" num=""><math display="inline"><mrow><mtext>n=0,1,2.....L/2-1</mtext></mrow></math><img file="EP0781052A2_D0008.tif" /></maths> but convolution of filter coefficients and reordered DCT coefficients may be performed in the DCT domain, which may be given by the following equation.<maths id="math0009" num="(5)"><math display="block"><mrow><msup><mrow><mtext>H</mtext></mrow><mrow><mtext>'</mtext></mrow></msup><mtext>(m) = 2</mtext><apply><sum /><lowlimit><mtext>n = 0</mtext></lowlimit><uplimit><mtext>N-1</mtext></uplimit><mrow><mtext>h'(n)cos(Πm(n+1/2)/N)</mtext></mrow></apply><mtext>   m = 0,1...N-1</mtext></mrow></math><img file="EP0781052A2_D0009.tif" /></maths>

0043Up-sampling and Down-sampling block 304 may be used to interpolate a vector <maths id="math0010" num=""><math display="inline"><mrow><mtext>k(n) = [x(n)]</mtext></mrow></math><img file="EP0781052A2_D0010.tif" /></maths> to twice its original size by upsampling, where x(n) is the individual vector in R. Upsampling may involve a function f(n) with vector k(n) padded with zeros on the right.<maths id="math0011" num=""><math display="block"><mrow><mtext>f(n) = {x1(n), x2(n), x3(n), x4(n), x5(n), x6(n), x7(n), x8(n), 0, 0, 0, 0, 0, 0, 0, 0};</mtext></mrow></math><img file="EP0781052A2_D0011.tif" /></maths> Assuming k(n) to be a vector at positions 1,2,3,4,5,6,7,8, computing C_IIE, is represented by eq(4). The upsampling property states:<maths id="math0012" num="(6)"><math display="block"><mrow><msub><mrow><mtext>C</mtext></mrow><mrow><mtext>IIE</mtext></mrow></msub><msub><mrow><mtext> = X</mtext></mrow><mrow><mtext>u</mtext></mrow></msub><msub><mrow><mtext>(m) = (f(n)</mtext></mrow><mrow><mtext>ä</mtext></mrow></msub><msub><mrow><mtext>(m) - f(n)</mtext></mrow><mrow><mtext>ä</mtext></mrow></msub><mtext>(N-m)/√2   m = 0.1....N-1</mtext></mrow></math><img file="EP0781052A2_D0012.tif" /></maths>

0044Using eq(1), we compute<maths id="math0013" num="(7)"><math display="block"><mrow><msub><mrow><mtext>C</mtext></mrow><mrow><mtext>IIE</mtext></mrow></msub><msub><mrow><mtext>{X</mtext></mrow><mrow><mtext>u</mtext></mrow></msub><mtext>(m)}</mtext></mrow></math><img file="EP0781052A2_D0013.tif" /></maths>    where the values of<maths id="math0014" num=""><img file="EP0781052A2_D0014.tif" /></maths>

0045It may be possible to implement digital filters in the DCT domain in Anti-aliasing Anti-imaging block 303. The convolution form of the Remez Exchange low-pass filter (eq(4)) in the DCT domain may be given as follows:<maths id="math0015" num="(8)"><math display="block"><mrow><msup><mrow><mtext>H</mtext></mrow><mrow><mtext>'</mtext></mrow></msup><mtext>(m) = 2</mtext><apply><sum /><lowlimit><mtext>n = 0</mtext></lowlimit><uplimit><mtext>N-1</mtext></uplimit><mrow><msup><mrow><mtext>h</mtext></mrow><mrow><mtext>'</mtext></mrow></msup><mtext>(n)cos(Πm(n+1/2/N)</mtext></mrow></apply><mtext>   m = 0,1...N-1</mtext></mrow></math><img file="EP0781052A2_D0015.tif" /></maths>

0046Using upsampled pixels computed from eq(7) and low-pass filter coefficients in the DCT domain from eq(8) processed in Up-sampling and Down-sampling block 304, a resized (interpolated) signal is computed using the following equation.<maths id="math0016" num=""><math display="block"><mrow><mtable><mtr><mtd><mrow><mtable><mtr><mtd><mrow><mtable><mlabeledtr><mtext>(9)</mtext><mtd><mrow><msub><mrow><mtext>V</mtext></mrow><mrow><mtext>1</mtext></mrow></msub><msub><mrow><mtext>(m) = C</mtext></mrow><mrow><mtext>IE</mtext></mrow></msub><mtext>{K(n)v(n)}   m,n = 0,1,2...N</mtext></mrow></mtd></mlabeledtr></mtable></mrow></mtd></mtr><mtr><mtd><mrow><mtable><mlabeledtr><mtext>(10)</mtext><mtd><mrow><msup><mrow><mtext>y(n) = C</mtext></mrow><mrow><mtext>-1</mtext></mrow></msup><msub><mrow><mtext> </mtext></mrow><mrow><mtext>IIE</mtext></mrow></msub><msub><mrow><mtext>(C</mtext></mrow><mrow><mtext>IIE</mtext></mrow></msub><msup><mrow><mtext>{2</mtext></mrow><mrow><mtext>±</mtext></mrow></msup><msup><mrow><mtext>h</mtext></mrow><mrow><mtext>1</mtext></mrow></msup><msup><mrow><mtext>(n)</mtext></mrow><mrow><mtext>±</mtext></mrow></msup><mtext>V,(m)})   m,n = 0,1,2...N </mtext></mrow></mtd></mlabeledtr></mtable></mrow></mtd></mtr></mtable></mrow></mtd></mtr></mtable></mrow></math><img file="EP0781052A2_D0016.tif" /></maths>

0047The type-2 IDCT may be given by the following equation<maths id="math0017" num="(11)"><math display="block"><mrow><msup><mrow><mtext>(x(n) = (2/N)</mtext></mrow><mrow><mtext>1/2</mtext></mrow></msup><apply><sum /><lowlimit><mtext>m = 0</mtext></lowlimit><uplimit><mtext>N-1</mtext></uplimit><mrow><msub><mrow><mtext>k</mtext></mrow><mrow><mtext>m</mtext></mrow></msub><msup><mrow><mtext>X</mtext></mrow><mrow><mtext>(2)</mtext></mrow></msup><msup><mrow><mtext>(m)cos[(2n+1)mΠ]/2</mtext></mrow><mrow><mtext>?</mtext></mrow></msup><mtext>N</mtext></mrow></apply><mtext>   n = 0,1...N-1</mtext></mrow></math><img file="EP0781052A2_D0017.tif" /></maths> where <ul id="ul0002" list-style="none" compact="compact"><li>X⁰⁽²⁾(m) is the reordered DCT pixel coefficients.</li></ul>

0048The following steps may be used to interpolate an image by two in both dimensions in Up-sampling and Down-sampling block 304: <ul id="ul0003" list-style="none" compact="compact"><li>1) Type (2) forward 2-dimensional DCT of an image by block NxN using eq(5).</li><li>2) Upsample pixels using eq(6) and eq(7), which may result in a block of size 2N x 2N still in the DCT domain.</li><li>3) Multiply each block 2N x 2N block with 2-D DCT domain low-pass filter coefficients (eq(8)) in Anti-aliasing Anti-imaging block 303 to accomplish anti-imaging low-pass filtering.</li><li>4) Perform Type-2 Inverse IDCT on each resulting block of step-3 in IDCT block 305 which may result in an image doubled in both directions using eq(11) which may then be output to Motion Compensation block 307.</li></ul>

0049The following example illustrates upsampling, convolution-multiplication of a motion video image DCT in one-dimension, and enlarging to an SIF 352x240 image in Decoder 300 using Up-sampling and Down-sampling block 304, Anti-aliasing Anti-imaging block 303, IDCT block 305, and programmable coefficients 306 in the preferred embodiment of the present invention. <ul id="ul0004" list-style="none" compact="compact"><li>1) Take the following 8 signals:    x(n) = {1,2,3,4,5,6,7,8} at positions 1,2,3,4,5,6,7,8</li><li>2) Compute X_II(m) using eq(2):    X2(m) : 1.. 8 <dl id="dl0001" compact="compact"><dt>X2(m) =</dt><dd>12.7279 -6.4423 0.0000 -0.6735 0.0000 -0.2009 0.0000 -0.0507</dd></dl></li><li>3) Upsample in Up-sampling Down-sampling block 304 by padding X2(m), with 8 zeros.    X2(m) for 9 through 16 <dl id="dl0002" compact="compact"><dt>X2(m) =</dt><dd>0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000</dd></dl> x(n) may now be twice its original length.</li><li>4) Compute C_IE using eq(6). <dl id="dl0003" compact="compact"><dt>C_IE =</dt><dd>9.0000 -4.5554 0.0000 -0.4762 0.0000 -0.1421 0.0000 -0.0359 0.0000 0.0359 0.0000 0.1421 0.0000 0.4762 0.0000 4.5554</dd></dl> Except for the first term, which is a DC coefficient, the even term is always a zero. </li><li>5) Multiply upsampled pixels obtained in step 4 and Remez low-pass filter coefficients given in table 1 in the DCT domain using eq(8).</li><li>6) Compute the 2D IDCT on the product obtained in step 4, using eq(1). The result is an interpolated signal rx(n) which is twice the size of the original signal x(n).</li></ul>

0050Using the same basic principle to interpolate a 352x240 SIF frame, interpolate the frame in both directions. The result may be an SIF frame of size 704x480. After upsampling the 8x8 block, the result is a 16x16 block in the DCT domain with padded zeros. After multiplying the reordered DCT coefficients with the low-pass filter coefficients in DCT domain and computing the inverse type-2 DCT of the product, an SIF image frame whose size is 704x480 may be the result.

0051Interpolating the old 352x240 SIF image frame in both directions in the spatial domain with seven tap filter coefficients [-29 0 140 256 140 0 -29], still results in a 704x480 interpolated image frame.

0052Examining the 704x480 interpolated image frame in spatial frequency domain and the DCT domain, the DCT domain image appeared to retain more information than the spatial frequency domain interpolation. Such a result may be due to: <ul id="ul0005" list-style="none"><li>1) Losslessness in the DCT domain relative to the spatial frequency domain.</li><li>2) The DCT domain resizing being a point-wise convolution-multiplication and sampling rate change, hence filter coefficients should be those of an even-tap filter. For symmetric convolution, the maximum number of filter coefficients may be twice the DCT block size resulting in a 32 tap filter. Filters with many taps result in sharper frequency response. Using a 32 tap filter may not affect the number of operations to achieve a Remez exchange response. Consequently, for DCT based interpolation, the longest possible tap filter may be used, without any extra hardware or latency involved resulting in a better response than that of the 7-tap spatial frequency domain interpolation.</li></ul>

0053The following data may represent an approximate comparison for a 32-tap DCT domain interpolation and a 7-tap spatial frequency domain interpolation: <ul id="ul0006" list-style="none"><li>1) Spatial frequency domain interpolation: Interpolating from YUV 4:2:0 to YUV 4:2:2 format on a 4:2:0 SIF picture, whose chrominance size is 176x120, averaging the pixels using a 7-tap filter whose coefficients are [-29 0 140 256 140 0 -29] requires several steps. Assuming 3 multiplication operations which comprise 3 shift operations and 2 addition operations each, and 2 addition operations, 17 total operations are required. To interpolate 176x120 = 21120 pixels, it takes 359040 basic operations per chrominance component in the spatial frequency domain. For both chrominance components, interpolation in spatial frequency domain for 4:2:0 to 4:2:2 format takes approximately 720000 operations. To interpolate from 4:2:0 to 4:4:4 in the spatial frequency domain, 4:2:0 to 4:2:2 interpolation may first be performed. Interpolation may then be performed from 4:2:2 to 4:4:4 using the same basic principle. For both chrominance components, the number of basic operations to interpolate 4:2:0 to 4:4:4 may be approximately 1.4 million. </li><li>2) Interpolation in DCT Domain: Type-2 IDCT may be performed on a 16x8 block to begin 4:2:0 to 4:2:2 interpolation. There may be 1724 basic operations per 16x8 block comprising 160 multiplies and 864 adds, assuming 4 shifts and 3 adds per multiply operation. For each chrominance component with a size of 176x120 there may be 330 blocks resulting in 570000 basic operation per chrominance component per interpolated frame. For both chrominance components, 1040000 operations may be required. </li></ul>

0054The IDCT block size may change from 16x8 to 16x16 for interpolation between 4:2:0 to 4:4:4 and the number of operations may double. The number of operations to interpolate both the chrominance components may be 2.6 million.

0055For encoded bit-streams in YUV 4:2:2 or YUV 4:4:4 formats, no modifications to the hardware decoder of the present invention may be required to perform decimation. Decimation may involve downscaling of images by simply eliminating or throwing away information from the image. Decimation in the DCT domain may involve the same basic principles used to interpolate in Up-sampling Down-sampling block 304.

0056Decimating a sequence x(n) to half its original size may first require an anti-aliasing low-pass filter step which may be performed in Anti-aliasing Anti-imaging block 303, then downsampling pixel coefficients are generated in Up-sampling and Down-sampling block 304. Finally, the type-2 IDCT may be computed in IDCT block 305.

0057In the preferred embodiment of the present invention, a Remez Exchange low-pass filter may be used to perform anti-aliasing low-pass filtering associated with decimation in Anti-aliasing Anti-imaging block 303. The pass-band and stop-band calculations for the anti-aliasing low-pass filter for decimation purposes and the low-pass filter for interpolation may be identical as illustrated in Figure 4.

0058If X(m) represents the inverse quantized DCT coefficients of the image frame to be downscaled, the coefficients may be low-pass filtered in the DCT domain using the following equation:<maths id="math0018" num=""><img file="EP0781052A2_D0018.tif" /></maths>

0059Y(m) represents the low-pass filtered sequence of X(m).

0060To perform downsampling, Y(m) may be rearranged used directly to compute the type-2 Inverse DCT of the image frame pixels.

0061The filtered coefficients are rearranged as follows:<maths id="math0019" num="(13)"><math display="block"><mrow><msub><mrow><mtext>Yd(m) = (Y</mtext></mrow><mrow><mtext>II</mtext></mrow></msub><msub><mrow><mtext>(m) - Y</mtext></mrow><mrow><mtext>II</mtext></mrow></msub><mtext>(N-m))/√2   m = 0,1...N-1</mtext></mrow></math><img file="EP0781052A2_D0019.tif" /></maths>

0062The type-2 IDCT is performed on Yd(m) using eq(11), resulting in a decimated image frame.

0063The following steps may be used to perform decimation in both directions on an image frame. <ul id="ul0007" list-style="none" compact="compact"><li>1) DCT of N x N (8 x *) block of the image frame using eq(2).</li><li>2) Convolution of transformed coefficients and low-pass filter coefficients in the DCT domain.</li><li>3) Rearranging DCT coefficients using eq(13).</li><li>4) Inverse IDCT using eq(11).</li></ul>

0064Figure 5 is a graph illustrating the frequency response of the Remez Exchange filter of Figure 4. Response curve 503 of graph 500 with a Remez Exchange characteristic shows desirable pass band characteristics that correlate to the pass-band and stop-band characteristics as illustrated in Figure 4. Pass band limit 501 clearly shows uniform characteristics in the Remez Exchange Filter pass band. Stop band limit 502 shows linear in-band characteristics and near zero energy at frequencies above stop band limit 502.

0065Figure 6 is a diagram illustrating an image processed using prior art spatial domain interpolation. The quality of edges and high spatial frequency information on image 600 may be compromised although not complete indiscernible.

0066Figure 7 is a diagram illustrating an image processed using DCT domain interpolation of the present invention. Image 700 was rendered using the same base image as in Figure 6. Image 700 may show superior edge rendering and more detail overall using DCT domain interpolation.

0067Figure 8 is a graph illustrating a signal filtered using the present invention. After reordering up or downsampled video stream data, the filtering process produces graph 800. Alternate points on graph 800 prior to filtering were zero points. Depending on number of taps used in filtering results of the DCT domain interpolation, graph 800 may become smoother in direct proportion to number of taps.

0068With rapid evolution of multimedia technology, there may be a need for higher resolution motion video. An MPEG decoder which takes advantage of DCT domain interpolation, allows for programmable color space format decoding and translation by changing IDCT coefficients and block size, and which allows for resizing may be advantageous. Even though the number of operations required for DCT domain transformations may be significantly greater than for spatial frequency domain transformations, costs associated with additional hardware required to perform interpolation in the spatial frequency domain far exceeds the cost of additional operations required by DCT domain interpolation with no additional hardware.

0069While the preferred embodiment and various alternative embodiments of the invention have been disclosed and described in detail herein, it may be apparent to those skilled in the art that various changes in form and detail may be made without departing from the spirit and scope of the invention.

0070For example, while the IDCT described herein may be type 2, the present invention could be practiced on systems using one of many forms of IDCT. The color space formats while encompassing those commonly known and used, could be practiced on any color space format. Moreover, although the preferred embodiment is drawn to an implementation in an integrated circuit, the present invention may be applied to a series of integrated circuits, a chipset, or in other circuitry within a computer system without departing from the spirit and scope of the present invention.