1. Technical Field of the Invention
The present invention relates to motion picture compression circuits for pictures such as television pictures, and more particularly to a compression circuit complying with H.261 and MPEG standards.
2. Description of Related Art
Intra (“I”) pictures are not coded with reference to any other pictures. Predicted (“P”) pictures are coded with reference to a past intra or past predicted picture. Bidirectional (“B”) pictures are coded with reference to both a past picture and a following picture.
The pixels of a 16 by 16-pixel square are arranged in a so-called “macroblock”. A macroblock includes four 8 by 8-pixel luminance blocks and two or four 8 by 8-pixel chrominance blocks. The processes hereinafter described are carried out by blocks of 8 by 8 pixels.
The blocks of each macroblock of picture 11 are submitted at 10 to a discrete cosine transform (DCT) followed at 11 by a quantization (Q). A DCT transforms a matrix of pixels (a block) into a matrix whose upper left corner coefficient tends to have a relatively high value. The other coefficients rapidly decrease as the position moves downwards to the right. Quantization involves dividing the coefficients of the matrix so transformed, such that a large number of coefficients which are a distance away from the upper left corner are cancelled.
At 12, the quantified matrices are subject to zigzag scanning (ZZ) and to run/level coding (RLC). Zigzag scanning has the consequence of improving the chances of consecutive series of zero coefficients, each of which is preceded by a non-zero coefficient. The run/level coding mainly includes replacing each series from the ZZ scanning with a pair of values, one representing the number of successive zero coefficients and the other representing the first following non-zero coefficient.
At 13, the pairs of values from the RLC are subject to variable length coding (VLC) that includes replacing the more frequent pairs with short codes and replacing the less frequent pairs with long codes, with the aid of correspondence tables defined by the H.261 and MPEG standards. The quantification coefficients can be varied from one block to the next by multiplication by a quantization coefficient. That quantization coefficient is inserted during variable length coding in headers preceding the compressed data corresponding to macroblocks.
Macroblocks of an intra picture are used to compress macroblocks of a subsequent picture of predicted or bidirectional type. Thus, decoding of a predicted or bidirectional picture is likely to be achieved from a previously decoded intra picture. This previously decoded intra picture does not exactly correspond to the actual picture initially received by the compression circuit, since this initial picture is altered by the quantification at 11. Thus, the compression of a predicted or intra picture is carried out from a reconstructed intra picture I1 rather than from the real intra picture I1, so that decoding is carried out under the same conditions as encoding.
The reconstructed intra picture I1r is stored in a memory area M2 and is obtained by subjecting the macroblocks provided by the quantification 11 to a reverse processing, that is, at 15 an inverse quantification (Q−1) followed at 16 by an inverse DCT (DCT−1).
The processing of the macroblocks of the predicted picture P4 is carried out from so-called predictor macroblocks of the reconstructed picture I1r. Each macroblock of picture P4 (reference macroblock) is subject to motion estimation (ME) at 17 (generally, the motion estimation is carried out only with the four luminance blocks of the reference macroblocks).
This motion estimation includes searching in a window of picture I1r for a macroblock that is nearest, or most similar to the reference macroblock. The nearest macroblock found in the window is the predictor macroblock. Its position is determined by a motion vector V provided by the motion estimation. The predictor macroblock is subtracted at 18 from the current reference macroblock. The resulting difference macroblock is subjected to the process described with relation to
Like the intra pictures, the predicted pictures serve to compress other predicted pictures and bidirectional pictures. For this purpose, the predicted picture P4 is reconstructed (P4r) in a memory area M3 by an inverse quantification at 15, inverse DCT at 19, and addition at 19 of the predictor macroblock that was subtracted at 18.
The vector V provided by the motion estimation 17 is inserted in a header preceding the data provided by the variable length coding of the currently processed macroblock.
At 20, the mean value of the two obtained predictor macroblocks is calculated and is subtracted at 18 from the currently processed macroblock.
The bidirectional picture is not reconstructed because it is not used to compress another picture.
The motion estimation 17 provides two vectors V1 and V2 indicating the respective positions of the two predictor macroblocks in pictures I1r and P4r with respect to the reference macroblock of the bidirectional picture. Vectors V1 and V2 are inserted in a header preceding the data provided by the variable length coding of the currently processed macroblock.
In a predicted picture, an attempt is made to find a predictor macroblock for each reference macroblock. However, in some cases, using the predictor macroblock that is found may provide a smaller compression rate than that obtained by using an unmoved predictor macroblock (zero motion vector), or even smaller than the simple intra processing of the reference macroblock. Thus, depending upon these cases, the reference macroblock is submitted to either predicted processing with the vector that is found, predicted processing with a zero vector, or intra processing.
In a bidirectional picture, an attempt is made to find two predictor macroblocks for each reference macroblock. For each of the two predictor macroblocks, the process providing the best compression rate is determined, as indicated above with respect to a predicted picture. Thus, depending on the result, the reference macroblock is submitted to either bidirectional processing with the two vectors, predicted processing with only one of the vectors, or intra processing.
Thus, a predicted picture and a bidirectional picture may contain macroblocks of different types. The type of a macroblock is also data inserted in a header during variable length coding. According to MPEG standards, the motion vectors can be defined with an accuracy of half a pixel. To search a predictor macroblock with a non integer vector, first the predictor macroblock determined by the integer part of this vector is fetched, then this macroblock is submitted to so-called “half-pixel filtering”, which includes averaging the macroblock and the same macroblock shifted down and/or to the right by one pixel, depending on the integer or non-integer values of the two components of the vector. According to H.261 standards, the predictor macroblocks may be subjected to low-pass filtering. For this purpose, information is provided with the vector, indicating whether filtering has to be carried out or not.
The succession of types (intra, predicted, bidirectional) is assigned to the pictures in a predetermined way, in a so-called group of pictures (GOP). A GOP generally begins with an intra picture. It is usual, in a GOP, to have a periodical series, starting from the second picture, including several successive bidirectional pictures, followed by a predicted picture, for example of the form IBBPBBPBB . . . where I is an intra picture, B a bidirectional picture, and P a predicted picture. The processing of each bidirectional picture B is carried out from macroblocks of the previous intra or predicted picture and from macroblocks of the next predicted picture.
The various functional blocks that are used in a typical prior art functional implementation are shown in
In
The quantized macroblock is forwarded to an encoding process and an inverse quantization block. The encoding process takes the quantized macroblock and zig-zag encodes it, performs run level coding on the resultant data, then variable length packs the result, outputting the now encoded bitstream.
The bitstream is monitored and can be controlled via feedback to a rate control system. This controls quantization (and dequantization) to meet certain objectives for bitstream. A typical objective is a maximum bit-rate, although other factors can also be used.
The inverse quantization block in
The various blocks required to generate the encoded output stream have different computational requirements, which themselves can vary according to the particular application or user selected restrictions. Throttling of the output bitstream to meet bandwidth requirements is typically handled by manipulating the quantization step.
Pure hardware architectures, while potentially the most efficient, suffer from lack of flexibility since they can support only a restricted range of standards; moreover they have long design/verification cycles. On the other hand, pure software solutions, while being the most flexible, require high-performance processors unsuited to low-cost consumer applications.
It would be desirable to provide an architecture that allowed for relatively flexible bitstream control while reducing the amount of software-based processing power required.
In an embodiment, a decoder circuit comprises: a processor configured to inverse quantize macroblocks to generate inverse quantized macroblocks; an inverse discrete cosine transformation circuit that processes the inverse quantized macroblocks from the processor to generate IDCT transformed macroblocks; and an addition circuit that adds a single IDCT transformed macroblock and a corresponding predictor macroblock to generate a reconstructed picture macroblock.
In an embodiment, a method for decoding an encoded bitstream, comprises: inverse quantizing decoded macroblocks in a processor to generate inverse quantized macroblocks; generating inverse discrete cosine transformation (IDCT) transformed macroblocks from the inverse quantized macroblocks; and adding a single IDCT transformed macroblock and a corresponding predictor macroblock to generate a reconstructed picture macroblock.
In an embodiment, a video compression circuit comprises: a discrete cosine transform (DCT) circuit for accepting prediction error macroblocks and generating DCT transformed macroblocks; a processor being configured to quantize the DCT transformed macroblocks to generate quantized macroblocks, and to inverse quantize the quantized macroblocks to generate inverse quantized macroblocks; an inverse discrete cosine transform (IDCT) circuit, wherein the IDCT circuit transforms the inverse quantized macroblocks to generate reconstructed prediction error macroblocks; and an addition circuit for adding a single reconstructed prediction error macroblock and a corresponding predictor macroblock to generate respective reconstructed macroblocks for use in the encoding of other macroblocks.
In an embodiment, a method of generating a compressed video bitstream comprises: generating DCT transformed macroblocks by applying prediction error macroblocks to a discrete cosine transform (DCT) circuit; quantizing the DCT transformed macroblocks to generate quantize macroblocks; inverse quantizing the quantize macroblocks to generate inverse quantize macroblocks; generating reconstructed prediction error macroblocks by applying the inverse quantize macroblocks to a IDCT circuit; and adding a single reconstructed prediction error macroblock and a corresponding predictor macroblock to generate respective reconstructed macroblocks for use in the encoding of other macroblocks.
In an embodiment, an encoder/decoder circuit comprises: a discrete cosine transform (DCT) circuit to generate DCT transformed macroblocks from prediction error macroblocks; a processor configured to quantize the DCT transformed macroblocks to generate quantized macroblocks, and to inverse quantize the quantized macroblocks to generate inverse quantized macroblocks; an inverse discrete cosine transform (IDCT) circuit to transform the inverse quantized macroblocks to generate reconstructed prediction error macroblocks; an addition circuit to add a single reconstructed prediction error macroblock and a corresponding predictor macroblock to generate respective reconstructed macroblocks; and a control circuit to configure the encoder/decoder circuit to encode or decode a bitstream.
In an embodiment, a method for encoding and decoding in an encoder/decoder circuit having a control circuit to configure the encoder/decoder circuit for encoding or decoding mode comprises: generating DCT transformed macroblocks by applying prediction error macroblocks to a discrete cosine transform (DCT) circuit; quantizing the DCT transformed macroblocks to generate quantized macroblocks; inverse quantizing the quantized macroblocks to generate inverse quantized macroblocks; generating reconstructed prediction error macroblocks by applying the inverse quantized macroblocks to the IDCT circuit; and adding the reconstructed prediction error macroblocks and corresponding predictor macroblocks to generate respective reconstructed macroblocks; wherein the reconstructed macroblocks are useful either as decoded reconstructed picture macroblocks or for encoding other macroblocks.
A more complete understanding of the method and apparatus may be acquired by reference to the following Detailed Description when taken in conjunction with the accompanying Drawings wherein:
The functional blocks include a subtraction circuit 300 for subtracting each predictor macroblock, as supplied by the motion estimation engine (described later) from its corresponding picture macroblock, to generate a prediction error macroblock. For an I picture, there is no predictor, so the macroblock is passed through the subtraction circuit with no change.
The prediction error macroblock is supplied to a DCT circuit 301 where a forward discrete cosine transform (DCT) is performed. Such hardware and its operation are well known in the prior art and will not be described here in further detail.
The output of the DCT is streamed to a processor 302 (described later) which performs the quantization, zig-zag coding, a run level coding steps in the encoding process. The resultant data is variable length coded and output as an encoded bitstream. In the simplified schematic of
The processor 302 also performs inverse quantization (Q−1), and the resultant inverse quantized macroblocks are sent to an inverse DCT (IDCT) circuit 303 via a streaming interface. An inverse DCT (IDCT) is performed and the resultant reconstructed error macroblock is added to the original predictor macroblock (for P and B pictures only) by an addition circuit 304. The predictor macroblocks have been delayed in a delay buffer 305. For I and P pictures, the macroblock is fully reconstructed after the IDCT circuit. The resultant reconstructed macroblocks are then stored in memory for use by the motion estimation engine in generating predictors for future macroblocks. This is necessary because it is reconstructed macroblocks that a decoder will subsequently use to reconstruct the pictures.
The motion estimation engine 400 outputs the macroblocks, associated predictor macroblocks and vectors, and other information such as frame type and encoding modes, to DCT/IDCT circuitry via a direct link. Alternatively, this information can be transferred over a data bus. Data bus transfer principles are well known and so is not described in detail.
The DCT and IDCT steps are performed in a DCT/IDCT block 401, which includes combined DCT/IDCT circuitry 301/303 that is selectable to perform either operation on incoming data. The input is selected by way of a multiplexer 402, the operation of which will be described in greater detail below. The output of the multiplexer is supplied to the delay block 305 and the DCT/IDCT circuitry 301/303. Additional data supplied by the motion estimation engine 400, such as the motion vector(s), encoding decisions (intra/non-intra, MC/no MC, field/frame prediction, field/frame DCT) is routed past the delay and DCT/IDCT blocks to a first streaming data interface SDI 403.
The outputs of the delay block and the DCT/IDCT circuitry are supplied to an addition circuit 304, the output of which is sent to memory 450. The output of the DCT/IDCT block 301/303 is also supplied to the first SDI port 403.
The first SDI port 403 accepts data from the DCT/IDCT block 301/303 and the multiplexer 402 and converts it into a format suitable for streaming transmission to a corresponding second streaming SDI port 404. The streaming is controlled by a handshake arrangement between the respective SDI ports. The second streaming SDI port 404 takes the streaming data from the first SDI port 403 and converts it back into a format suitable for use within the processor 302.
Once the data has been transformed back into a synchronous format, the processor performs quantization 405, inverse quantization 406 and zig-zag/run level coding 407 as described previously. It will be appreciated that the particular implementations of these steps in software is not relevant, and so is not described in detail.
After inverse quantization, the macroblock is returned to a third SDI port 408, which operates in the same way as the first streaming port to convert and stream the data to a fourth SDI port 409, which converts the data for synchronous use and supplies it to the multiplexer 402.
The processor 302 outputs the run level coded data to a fifth SDI port 410, which in a similar fashion to the first and third SDI ports, formats the data for streaming transmission to a sixth SDI port 411, which in turn reformats the data into a synchronous format. The data is then variable length coded and packed in hardware VLC circuitry 412. The particular workings of the hardware VLC packing circuitry 412 are well known in the art, are not critical to understanding the present solution and so will not be described in detail. Indeed, as mentioned previously, the VLC operation can be performed in software by the processor, for a corresponding cost in processor cycles.
It will be appreciated that a number of control lines and ancillary detail has been omitted for clarity. For example, it is clear the multiplexer and DCT/IDCT block 301/303 need to be controlled to ensure that the correct data is being fed to the DCT/IDCT block and that the correct operation is being performed. For example, when the initial DCT operation 301 is being performed, the multiplexer 402 is controlled to provide data from the bus (supplied by the motion estimation engine) to the DCT/IDCT block 301/303, which is set to DCT mode. However, when performing the IDCT operation 303, the multiplexer 402 sends data from the fourth SDI port 409 to the DCT/IDCT block 301/303, which is set to IDCT mode.
Similarly, some support hardware that would exist in the actual implementation has been omitted. An obvious example is buffers on the various inputs and output. It would be usual in such circuitry to include FIFO buffers supporting the SDI ports to maximize throughput. For the purposes of clarity, such support hardware is not explicitly shown. However, it will be understood by those skilled in the art to be implicitly present in any practical application.
It will be appreciated that, in the encoding mode described above, the DCT and IDCT functions of the DCT/IDCT block 301/303 will be performed in an interleaved manner, with one or more DCT operations being interleaved with one or more IDCT operations, depending upon the order of I, P and B pictures being encoded.
With slight modifications to control software and circuitry, the encoding circuitry described above can perform decoding of an encoded MPEG stream. This is because the inverse quantization software and IDCT hardware are common to the encoding and decoding process. There are at least three ways this can be achieved:
Option 1. If it is only required to offload the IDCT processing from the processor, the dequantized coefficient blocks can be streamed from the processor to the IDCT/DCT block 301/303 via the third and fourth SDI ports 408 and 409. The results of the IDCT are then read back via the first and second SDI ports 403 and 404.
Option 2. Option 1 can be extended to allow more of the decoding load to be passed to the DCT/IDCT block 401. In particular, the predictor blocks are read into the delay buffer 305. The coefficient blocks are then read in via the same route by the DCT/IDCT block 301/303 (in IDCT 30 mode). After the IDCT has taken place, the predictor and IDCT processed macroblocks are combined by the addition circuitry 304 and written to system memory via the system data bus.
Option 3. In an alternative to option 2, the motion estimation block is configured to provide the predictor blocks to the delay buffer 305 via the multiplexer 402. The coefficient blocks are provided to the DCT/IDCT block 301/303 (in IDCT mode), and the remainder of the procedure is as per the second decoding arrangement.
Although preferred embodiments of the method and apparatus of the present invention have been illustrated in the accompanying Drawings and described in the foregoing Detailed Description, it will be understood that the invention is not limited to the embodiments disclosed, but is capable of numerous rearrangements, modifications and substitutions without departing from the spirit of the invention as set forth and defined by the following claims.
Number | Date | Country | Kind |
---|---|---|---|
02251932.6 | Mar 2002 | EP | regional |
This application is a divisional of U.S. Application for patent Ser. No. 10/391,442, filed Mar. 17, 2003, which claims priority from European Application for Patent No. 02251932.6 filed on Mar. 18, 2002, the disclosures of which are hereby incorporated by reference.
Number | Date | Country | |
---|---|---|---|
Parent | 10391442 | Mar 2003 | US |
Child | 12020668 | US |