Field of the Invention

This invention relates generally to transcoding of digitally encoded signals, and more particularly to transcoding of signals that are digitally encoded by predictive coders.

Background

A predictive waveform encoder is a device for compressing the amount of information in a waveform (e.g., speech, image or video) by removing the statistical redundancy among its neighboring samples using prediction methods. Several ITU-T Recommendations for speech coding (ITU-T stands for the

Telecommunication Standardization Sector of the International Telecommunication Union; ITU-T is formerly known as CCITT, or International Telegraph and Telephone Consultative Committee), have adopted predictive coding techniques (for example, differential pulse-code mudulation, or DPCM, is used in

Recommendation G.721 ). In these predictive speech coders, an original speech sample is predicted based on past speech samples, and the prediction error (the difference between the original and the predicted samples), instead of the original sample, is quantized, and then digitally encoded by a noiseless coder to a bit stream. Since the energy of the prediction error is, on average, much smaller than the original speech signal, a high compression ratio can generally be obtained.

Predictive coding methods have also been used for image and video compression. In these applications, the spatial correlation among neighboring pixels in an image and, in the case of video, the temporal correlation between successive images can be exploited.

Typical predictive coders perform the prediction based on a replica of the reconstructed waveform. This ensures that the quantization error does not accumulate during reconstruction. Although the prediction accuracy is reduced (for coarse quantization), overall compression performance is generally improved.

State-of-the-art digital video coding systems utilize transform coding for spatial compression and a form of predictive coding known as motion-compensated prediction (MCP) for temporal compression. Video compression techniques that have recently been adopted in international standards (e.g., the MPEG standard developed by the International Standards Organization's Motion Picture Experts Group (ISO's MPEG) and ITU- T's H.261 ), or others that are under consideration for future standards, all employ a so-called block-matching MCP technique. In this method, each image in a video sequence is partitioned into NxN blocks, called macro blocks (MB's), where N is a predetermined integer. For each MB, a replica of the previously decoded image is searched to find an NxN window that best resembles that MB, and the pixels in that window are used as a prediction for that MB. The prediction error is then encoded using a combination of transform coding and scalar quantization followed by variable-length noiseless encoding.

Transcoding will be required in many applications of compressed digital video. For example, in some instances, it may be desirable to change the rate of a digital video bit stream in the network. Alternatively, when constant bit-rate (CBR) video traffic is to be carried over a cell-relay or

Asynchronous Transfer Mode (ATM) network, it may be desirable to convert the CBR stream into a variable bit-rate (VBR) stream to save bandwidth through statistical multiplexing. Transcoding may also be required for conversion between two video compression formats. For example, it may

be necessary to convert an MPEG-encoded video bit stream into an H.261 bit stream, or vice versa. Another important application of transcoding is multipoint video conferencing; here, transcoding may be needed to implement video mixing for continuous presence multipoint bridging.

FIG. 1 , numeral 100, is a block diagram schematic of a predictive waveform encoder as is known in the art. A sequence of vectors consisting of a group of samples η taken from an original waveform are processed to generate a sequence of quantized vectors Yj, where i = 0, 1 is a time index indicating the order in which the input vectors are processed. The dimensionality L of the input vectors is arbitrary. In typical speech applications L = 1 , whereas in many video compression applications, L > 1.

The encoder operates iteratively such that: (1 ) a predictor unit (102) generates a prediction of the input vector x\ represented by the vector pi based on one or more past reconstructed vectors z , j < i, using a predetermined linear prediction operator P\; (2) the vector pi is subtracted from η at a first combiner (104) to obtain the prediction error vector ej = η - Pi, wherein the predictor Pj is typically chosen to minimize the average energy of the prediction error e\; (3) the prediction error vector e is transformed by a transformation unit (106) according to Ej = Aj[eι], where Aj[ ] represents a linear transformation; (4) the vector Ej is quantized using a quantizer Qj (108) to obtain the quantized vector Yj = Ej + Dj, where Dj is a quantization error vector, and the quantized vector Yj is encoded into a binary word using a noiseless encoding method (e.g., a Huffman code), and then it is transmitted or stored; (5) the quantized vector Yj is then inverse transformed at Inverse Transformation Unit A\- ¹ [ . .0) to find the vector y = Af ¹ [Yj], where Af ¹ [ ] is an inverse transformation (i.e., Af ¹ [Aj[x]] = x); and (6) the vector p,- is added

by a second combiner (1 12) to y; to obtain the reconstructed vector Zj = yj + pj, which is stored for use in later iterations.

In most applications, the transformation Aj is fixed a priori, i.e., is predetermined, whereas Qj and Pj are varied using preselected adaptation algorithms. In some applications, the transformation A* is not used; then Aj = I, where I is an LXL identity matrix. In so-called forward adaptation, the parameters of Qj , Pj and Aj are passed to the decoder as side information, while in so-called backward adaptation, Qj , Pj and Aj are determined at the decoder from previously received information, so no side information needs to be sent.

Given the information on Qj, Pj and Aj, a decoder can reconstruct the vector ZJ. The decoder (200) first recovers the quantized vectors {Yj} from the received bit stream by decoding the noiseless source code and then obtains ZJ. As shown in Fig. 2, numeral 200, (1 ) the quantized vector Yj is first inverse transformed using the inverse transformation unit Aj ^-1 (202) to obtain yj = Aj ^_1 [Yj]; (2) a predictor (206) obtains the prediction pj of the input vector π from one or more past reconstructed vectors ZJ , j < i , using the prediction operator Pj, as in the encoder; and (3) a combiner (204), operably coupled to the predictor (206) and to the transformation unit (Aj ^-1 ) (202) adds the vector pj to yj to obtain the reconstructed vector ZJ.

The reconstructed vector z; can be represented as ZJ = η + dj, where dj = Ap ¹ [Dj] is an inverse-transformed version of the quantization error vector Dj. In other words, ZJ differs from the original vector π only by dj = Aj ^-1 [Dj]. To obtain good performance, the transformation Aj is chosen such that the error Aj ^-1 [Dj], or an appropriately weighted version of it, is kept small.

A transcoder first recovers the sequence of quantized vectors {Yj } from the received bit stream by decoding the

noiseless source code, converts {Yj } into a sequence of transcoded vectors {Yj'}, and then generates a new bit stream representing {Yj'} using the noiseless source code. The transcoder has full knowledge of the operators Qj, Aj and Pj used at the original encoder and decoder, either a priori or through received side information.

In prior art "decode and re-encode" transcoding, a quantized vector Yj is first decoded using the decoder of Fig. 2 to obtain the reconstructed vector ZJ = η + dj and then ZJ is re-encoded using an encoder, possibly with a different quantizer Qj', a different predictor Pj' or even a different transformation Aj', to obtain the transcoded vector Yj'. The transcoded vector can be decoded by the decoder of Fig. 2 using Qj', Pj ¹ and Aj'. The final reconstructed vector Zj' can then be represented as ZJ' = η + dj + dj', where dj'=

(Aj')- ¹ [Qj'] is a transformed version of the quantization error introduced by the transcoder.

Although conceptually straightforward, the implementation of the decode and re-encode method can be quite costly because of its high computational and storage requirements. Thus, there is a need for an efficient transcoding device and method that can be implemented with low complexity.

Brief Descriptions of the Drawings

FIG. 1 is a block diagram schematic of a predictive waveform encoder as is known in the art.

FIG. 2 is a block diagram schematic of a decoder that typically operates in conjunction with an encoder of FIG. 1 , as is known in the art.

FIG. 3, numeral 300, is a general block diagram schematic of a system for transcoding.

FIG. 4 is a block diagram schematic of a device in accordance with the present invention.

FIG. 5 is a block diagram schematic of the device of FIG. 4 shown with greater particularity.

FIG. 6 is an alternative block diagram schematic of the device of FIG. 4 shown with greater particularity.

FIG. 7 shows a typical configuration of an H.261 video encoder, as is known in the art.

FIG. 8 illustrates the partition of an image into increasingly smaller blocks in the H.261 standard.

FIG. 9 shows the indexing of the coefficients in the transform domain for the H.261 standard.

FIG. 10 shows a block diagram schematic of one embodiment of the one-step transcoder of the present invention.

FIG. 11 is a block diagram schematic of an H.261 decoder as is known in the art.

FIG. 12 is a flow chart of one embodiment of steps in accordance with the method of the present invention.

Detailed Description of a Preferred Embodiment

FIG. 3, numeral 300, is a general block diagram schematic of a system for transcoding as is known in the art. The input of the transcoder (304) is a bit stream generated by a predictive waveform encoder (302) such as the H.261 video encoder. Its output is another bit stream which can be decoded by a predictive waveform decoder (306). The transcoder (304) modifies the bit stream according to a predetermined objective.

The transcoder device of the present invention, referred to herein as a "one-step transcoder," achieves the performance of the "decode and re-encode" transcoder with only two transformations and one prediction operation, provided that the predictor Pj' and the transformation Aj' used in the transcoder are the same as the operators Pj and Aj used in the original encoder, thus decreasing complexity in the transcoding operation. In addition, the one-step transcoder reduces the storage requirements.

In the one-step transcoder, shown in a block diagram schematic in Fig. 4, numeral 400, a modified version Sj of the quantization error vector is subtracted from the received quantized vector Yj , and the difference vector Ej' is re-quantized to obtain the transcoded vector Yj '. A quantization error calculator then computes the inverse-transformed quantization error vector dj' = Af ¹ [Dj'], where Dj' = Yj ' - Ej' is the quantization error vector and Aj" ¹ is an inverse transformation. A modifying circuitry determines the modified quantization error vector Sj based on past vectors dj', j < i.

The present invention includes a device for transcoding a sequence of quantized vectors Yj generated by a predictive waveform encoder utilizing modified quantization error vectors Sj. The device (400) includes an adder (402), a quantizer (404), a

quantization error vector calculator (QEVC) (406), and modifying circuitry (408). The adder (402) is operably coupled to receive at least a first quantized vector Yj and at least a first modified quantization error vector Sj and is utilized for generating a difference vector Ej' = Yj - Sj between the quantized vector Yj and the modified quantization error vector Sj. The quantizer (404) is operably coupled to the adder (402) and is used for quantizing the difference vector Ej' to obtain a transcoded vector Yj'. The quantization error vector calculator (QEVC) (406) is operably coupled to receive at least two of the vectors Yj, Ej' and Yj' and is utilized for generating the inverse-transformed error vector dj' = Ar ¹ [Dj'], where Dj' is a quantization error vector and Aj" ¹ is an inverse transformation. The modifying circuitry (408) operably couples the QEVC (406) to the adder (402) and is used for generating the modified quantization error vector Sj based on the past values of the vector dj'.

In an alternate embodiment, the transcoding device of the present invention may be selected to include an adder (402), a generalized quantizer (410), and modifying circuitry (408). In this implementation, the adder (402) is operably coupled to receive at least a first quantized vector Yj and at least a first modified quantization error vector Sj and is used for generating a difference vector Ej' = Yj - Sj between the quantized vector Yj and the modified quantization error vector SJ; the generalized quantizer (410) is operably coupled to the adder (402) and is used for quantizing the difference vector Ej' to obtain a transcoded vector Yj', and for receiving at least one of the vectors Yj and Ej' and generating an inverse-transformed error vector dj' = Af ¹ [Dj'], where Dj' is a quantization error vector and Aj- ¹ is an inverse transformation; and the modifying circuitry (408) is operably coupled to the generalized quantizer (410) and is used for generating the modified quantization error vector Sj based on past values of the vector dj'.

One embodiment of the one-step transcoder device of FIG. 4 is shown with greater particularity in Fig. 5, numeral 500. The device includes a first adder (502), a quantizer (504), a quantization error vector calculator (QEVC) (506) that includes a second adder (510) coupled to an inverse transformation unit (512), and modifying circuitry (508) that includes a predictor (514) coupled to a transformation unit (516). The first adder (502) and the quantizer (504) are coupled as shown for the adder (402) and the quantizer (404), respectively, of FIG. 4. The second adder (510) is operably coupled to receive vectors Ej' and Yj' and is utilized for generating the quantization error vector Dj'. Th e inverse transformation unit (512) is operably coupled to the second adder (510) and is utilized for generating the inverse- transformed error vector dj' = Aj- ¹ [Dj'], where Dj' is a quantization error vector and Aj- ¹ is an inverse transformation. The predictor

(514) of the modifying circuitry (508) is operably coupled to the inverse transformation unit (512) of the QEVC (506) and generates the predicted quantization error vector Sj. The transformation unit (516) is operably coupled to the predictor (514) and is used for transforming the predicted quantization error vector SJ to vector Sj as described more fully below and providing the modified quantization error vector Sj to the first adder (502) based on past values of the vector dj'.

Past inverse-transformed quantization error vectors dj' - Aj ^-1 [Dj'], j < i, are passed through the prediction operator Pj to obtain the the predicted quantization error vector SJ. The vector SJ is transformed again to obtain the modified quantization error vector Sj = AJ[SJ]. Then the vector Sj is subtracted from the received quantized vector Yj to obtain the error vector Ej' = Yj - Sj.

The error vector Ej' is quantized using the quantizer Qj' (504) to obtain the transcoded vector Yj' = Ej' + Dj', where Dj' is the quantization error vector introduced by the transcoder. The transformed quantization error vector dj' is obtained by first subtracting the vector Ej' from Yj' (using the second adder (510))

to obtain the quantization error vector Dj' = Ej' + Yj', and then transforming Dj' (using the inverse transformation unit (512)) to obtain dj' = Aj- ¹ [Dj'].

The one-step transcoder generates the same transcoded sequence {Yj'} as the decode and re-encode transcoder. This can be shown by proving that the signal at the input of the quantizer is the same in both cases:

First consider the decode and re-encode transcoder. In this case, the input to the quantizer in the re-encoder can be written as

= Zj - Aj[r _pi + dpi + dpi'],

where Zj = AJ[ZJ] is a transformed version of the decoder output z\ (in the transcoder) and r j, d _P j and d _P j' represent the outputs of the predictor Pj at time i if excited individually by the sequences {n}, {dj} and {dj'}, respectively.

Similarly, the input of the quantizer in the one-step transcoder can be written as

= Zj - Aj[r _pi + dpi] - Aj[d _P i'].

Since Aj is a linear operator, it follows that Bj = Cj.

The present invention relies on the fact that the quantized vector Yj can be re-quantized without any error accumulation in the reconstruction, provided that a modified quantization error vector Sj is added to Yj before re-quantization. This compensates for the transcoder quantization error added by the prediction loop in the reconstruction.

The one-step transcoder of the present invention can also be implemented in other ways. One alternative structure is shown in Fig. 6.

FIG. 6, numeral 600, sets forth a block diagram schematic of an implementation of a device in accordance with the present invention wherein the difference between the input Yj and the output Yj' is fed through the feedback loop. The quantization error vector calculator (606) includes a second adder (610) that is operably coupled to receive Yj and Yj' for obtaining said difference {Xj}, an inverse transformation unit (612) that is operably coupled to the second adder (610) for providing vector XJ, and a third adder (614) for receiving and combining XJ and SJ, and modifying circuitry (608) that includes a predictor (616) that is operably coupled to the third adder (614), for utilizing dj' to provide the predicted quantization error vector SJ, and a tranformation unit (618) that is operably coupled to the predictor (616), for generating modified quantization error vector Sj. The difference in the implementation of FIG. 6 lies in the way the transformed quantization error vector dj' is generated:

The input vector Yj is subtracted from the transcoded vector Yj' to obtain the vector Xj = Yj' - Yj, which is then transformed to find XJ = Aj ^-1 [Xj']. The predicted quantization error vector SJ is added to XJ to determine dj' = XJ + SJ.

That the inverse-transformed quantization vector dj' is the same in both structures can be shown by noting that the vector XJ in FIG. 6 can be written as

xi - A|-1 [Y|' - Ei'] - si'.

Therefore, the implementations of FIGs. 5 and 6 provide the same performance.

The one-step transcoder may be utilized to change the rate of a bit stream generated by a video encoder operating according to the ITU-T Recommendation H.261. First, a typical configuration of an H.261 encoder is shown in FIG. 7, numeral 700.

The input to the encoder consists of a sequence of images scanned progressively at a nominal image rate of about 30 images per second. Each image consists of a luminance component Y and two color difference components C _B and C _R , sampled according to one of two formats, CIF (802) and QCIF (804), wherein:

CIF: 352 x 288 (Y), 176 x 144 (CR), 176 x 144 (CB) QCIF: 176 x 144 (Y), 88 x 72 (CR), 88 x 72 (CB).

Each image is partitioned into increasingly smaller segments as illustrated in FIG. 8, numeral 800. CIF images are divided into 12 Group Of Blocks (GOB's) and QCIF images are divided into three GOB's. Each GOB (806) consists of 33 macro blocks (MB's), and a MB consists of four luminance blocks (808) and two color difference blocks (810, 812), where each block has 64 pixels (814) arranged on an 8x8 grid. Each pixel is represented by an integer between 1 and 254.

MB's are the fundamental coding elements in H.26. The six

8x8 blocks in a MB are numbered from 1 to 6 as shown in FIG. 8. Let π _f k-m,n represent the pixel in position (m,n) in the k'th block of the i'th MB, where i = 0, 1 k = 1 6 and m, n = 0, 1 ,...,7.

Then the input vector η for the i'th MB can be represented as:

π = [π.1 ,0,0 n, 1 ,7,7, n, 2, o,o n,2,7,7, n, 3.0,0 n, 3,7,7. π.4,0,0 n,4,7,7. n,5,0,0 - ^" 1,5,7,7, n,6,0,0, n, 6,7,7] -

The operation of the encoder in FIG. 7, numeral 700, for the i'th MB is described as follows: First, a motion estimation unit

(702) utilizes an algorithm to determine a motion vector mi = (mji , rτij2). Typically, the algorithm searches the luminance pixels in the previous reconstructed image (stored in a frame buffer (704)) to find a 16x16 window Wj for which the "distance" between the pixels in that window and the corresponding pixels in the current MB is minimum. The motion vector mj represents the spatial offset between the window Wj and the current (i'th) MB.

The pixels that lie in the window Wj form the motion- compensated vector uι = [ui,ι ,o,o Ui.6,7,7] that is stored in the motion compensation unit (706). Thus, the motion estimation unit (702) is operably coupled to the frame buffer (704), and provides a motion compensated vector to the motion compensation unit (706), which is also operably coupled to the frame buffer (704). The mode unit (710) that is operably coupled to receive the input vector η determines the encoding mode

(Inter/lntra?). The prediction pi = [pi, 1 ,0,0, pι,6,7,7] of the input vector η is obtained from uι based on the encoding mode of the current MB:

a. In the intra mode, set pj = 0.

b. In the inter (predictive) mode:

b1. If the loop filter (708) is "out," set pi = UJ. b2. If the loop filter (708) is "in," filter the elements of uι (block-by-block) using a separable, two-dimensional, 3-tap FIR filter, and set the output of the loop filter equal to Pi.

in each case, the loop filter (708) is operably coupled to the motion compensation unit (706) to operate as set forth above.

At the first adder (712), the vector pj is subtracted from the input η to obtain the prediction error vector ej = [ej,ι ,o,o,

ej,6,7,7] = π - Pi- The vector ej is transformed to find Ej =

[Ej, 1 ,0,0, Ej, 6,7,7] = A[ej], where A[ ] here represents the

Discrete Cosine Transform (DCT) at a DCT unit (714) that is operably coupled to the first adder (712). The DCT is applied independently to each 8x8 block in the MB to obtain the transform coefficients Ej,k,s,t according to:

Ej, _k , _s ,t = 0.25 C(s) C(t) ∑∑ o<m,n<7 π, _k , _m , _n cos[π(2m+1 )s/16] cos[π(2n + 1 )t/1 6] ,

where C(s) = 1/V2 for s = 0, and 1 otherwise, and C(t) = 1/V2 for t = 0, and 1 otherwise. Here s and t are the transform domain variables. Note that the same transformation A[ ] is used in every MB. FIG. 9, numeral 900, shows how the indices t (902) and s (904), both in the range 0 to 7, are used in the transform domain.

A quantizer (716) is operably coupled to receive the transform coefficients Ej, ,s,t and quantizes the coefficients using a scalar quantizer which is uniform with step size Δj except for a dead-zone around 0. The reconstruction values of the quantizer are {0, ±aj, ±(aj + Δj), ±(aj + 2Δj), -2048 < ±(aj + 126Δj) < 2048}, where Δj = 2, 4 62 and a; = 3Δj/2 when Δj is odd and aj = 3Δj/2 -

1 , otherwise. The same step size Δj is utilized for all transform coefficients in the MB, except in the intra mode, the sample Ej,k,o,o is quantized using a uniform scalar quantizer of step size Δj = 8 with no dead-zone (also, since Ej,k,o,o ≥ 0, only the positive reconstruction values are needed in this case). The decision regions of the quantizer are selected to improve the image quality as much as possible.

The output of the quantizer (716) is the quantized transform vector Yj = [Yj,ι ,o,o Yi.6,7,7] = Ej + Dj, where Dj = [Dj,ι , ₀ ,o,

Di,6,7,7] is the quantization error vector.

The quantized vector Yj is input into an inverse DCT unit (724) and is further transformed to generate the vector yj =

[yi,ι ,o,o , yi.6,7,7] = A ^"1 [Yj], where A ^"1 [ ] is an inverse DCT. The pixels yj,k,m,n are determined according to:

yi,k-m,n = 0.25 ∑∑ ₀ <s,t<7 C(s) C(t) Yi.k.s.t cos[π(2m+1 )s/16] cos[π(2n+1 )t/1 6].

The vector pi is input to a second adder (726) and is added to yj to obtain the reconstructed vector z = [ZJ,I ,O,O ^z i,6,7,7] = yi

+ pj , and the pixels Zj,k,m,n are stored in the frame buffer (704).

The quantized transform coefficients Yj,k,s,t are typically encoded into a CBR bit stream, for example, by utilizing a variable-length encoder (720) with an output buffer (722) and then transmitted (or stored). First the coefficients in each block are converted from the 8x8 matrix format into a serial format using what is known as 'zig-zag scanning' (see FIG. 9), and then the coefficients in each block are represented by a sequence of (Run, Level) values where "Run" represents the number of zeroes before the next non-zero value "Level." These (Run, Level) values are then encoded using a binary variable-length code. The output of the variable-length encoder is typically buffered (720) in order to generate a CBR bit stream, and the quantization step size is adjusted by a quantizer control unit (718) to prevent buffer overflows (or underflows). The quantizer control unit (718) is operably coupled to the buffer (722) and provides an adjusting signal to the quantizer (716).

In addition to the quantized transform coefficients Yj,k,s,t, the encoder also transmits side information to allow the decoder to correctly reconstruct the coded signal. Side information includes the source format (CIF/QCIF), quantizer step size Δj, inter/intra decision, motion vector mi (in inter mode only) and the loop filter in/out (when motion vector is present).

The step size Δj can be kept fixed for an entire GOB. In that case, only one step size value per GOB needs to be transmitted as side information. It is also possible to change Δj inside the GOB. This provides a finer adjustment of the step size at the expense of a larger overhead.

The H.261 encoder also transmits side information to allow the encoder skip a block or a MB. For example, when all the coefficients Yj,k,s,t in an 8x8 block are zero, the encoder does not code these blocks at all. Similarly, when there is little motion, or when the motion estimation is nearly perfect, all coefficients in a MB may be zero. In that case, the encoder may skip the entire MB. When a block or a MB is skipped in the encoder, the decoder simply substitutes zeroes for the missing coefficients.

The maximum image rate in H.261 is approximately 30 images/sec, but the Recommendation allows the encoder to regularly skip 0, 1 , 2 or 3 images to achieve effective image rates of 30, 15, 10 and 7.5 images/sec. The encoder may also skip an image occasionally. This can be used, for example, immediately after encoding an image in the intra mode. Since the intra mode typically generates a large number of bits, skipping an image can help reduce the buffering delay.

The rate of a bit stream generated by an H.261 encoder can be changed using the transcoder of the present invention. A transcoder first decodes the received bit stream using a decoder for the variable-length code to obtain the sequence of (Run, Level) values, and then recovers the sequence of quantized vectors Yj =

[Yj, ι ,o,o Yj, 6,7,7] generated by the encoder. The decoder also recovers all side information. If a block or MB is not coded, the decoder inserts zeroes for the corresponding missing coefficients.

Typical subsequent operations of the one-step transcoder are described below (see FIG. 10, numeral 1000). A variable-length decoder (1002) outputs vector Yj to a first adder (1004) and provides framing information, inter/intra information, information about which blocks are coded (coded block pattern or CBP), step size Δj, motion vectors mj and loop filter information to various elements of the transcoder along a feedforward path. The first adder (1004) combines vector Yj and a modified quantization error vector Sj to provide vector Ej' to a quantizer (1006) and to a second adder (1014). The quantizer (1006) is operably coupled to the first adder (1004) and provides a quantized output vector Yj' to a variable-length encoder (1010) and to the second adder (1014). The variable length encoder (1010) is operably coupled to the quantizer (1006) and to receive information from the variable-length decoder (1002) and generates the output bits. The buffer (1012) is operably coupled to the variable-length encoder (1010) and provides a means for storing output bits before transmission and also provides an input to a quantizer control (1008). The quantizer control (1008) is operably coupled to the buffer (1012) and provides a control signal to the quantizer (1006) as described more fully above. The second adder (1014) is operably coupled to receive the vectors Ej* and Yj' and provides Dj'= Ej' - Yj'. An inverse DCT unit (1016) is operably coupled to the second adder (1014) and provides an inverse transform vector dj'. The frame buffer (1018) is operably coupled to the inverse DCT unit (1016) and provides output vector ZJ. A motion compensation unit (1020) is operably coupled to the frame buffer and to receive a motion vector from the variable- length decoder (1002) and is utilized to provide an output vector WJ. The loop filter (1022) is operably coupled to the motion compensation unit (1020), receives in/out loop filter information from the variable-length decoder (1002), and outputs the predicted quantization error vector SJ. A DCT unit (1024) is operably coupled to the loop filter (1022) and outputs vector Sj.

The above-cited vectors are further described as follows: ( 1 ) No motion estimation is performed in the transcoder. Instead, the motion vector mj = (rriji , rτij2) received from the encoder is used to determine the 16x16 window Wj, and the pixels in the transformed quantization error buffer (see below) that lie in that window are used to form the vector WJ = [WJ,I ,O,O

Wj,6,7,7]- Again, the motion vector mj represents the spatial offset between the window Wj and the current (i'th) MB.

The predicted quantization error vector Sj = [SJ,I ,O,O

Sj, 6, 7, 7] is obtained from WJ based on the encoding mode of the current MB:

a. In the intra mode, set SJ = 0.

b. In the inter (predictive) mode:

b1. If the loop filter is out, set SJ = WJ. b2. If the loop filter is in, filter the elements of WJ (block-by-block) using a separable, two-dimensional 3-tap FIR filter, and set the output of the filter equal to Sj.

(2) The vector Sj is transformed again to obtain the modified quantization error vector Sj = A[SJ], where A[ ] represents the DCT operation, according to:

Si, _k ,s,t = 0.25 C(s) C(t) ∑∑ o<m,n<7 Sj _fk ,m,n cos[π(2m+1 )s/16] cos[π(2n+1 )t/16] .

(3) The modified quantization error vector Sj =

[Sj, ,0,0 Sj,6,7,7] is subtracted from the coded vector Yj to obtain the error vector Ej' = [E'i,ι _f o,o E'i.6,7,7] = Yi - Sj.

(4) The coefficients E'j,k,s,t are (re)-quantized using an H.261 quantizer as in the encoder, possibly with a different step

size Δj', except for the term E'j,k,o,o in the intra mode the same step size is used: Δj = Δj' = 8. The result is the transcoded transform vector Yj' = [Y'j,ι ,o,o,---., Y'i.6,7,7] = Ej'+ Dj', where Dj' =

[D'i, ι ,o,o, D'j,6,7,7] is the quantization noise vector for the transcoder.

(5) The error vector Ej' is subtracted from Yj' to obtain the quantization error vector Dj', and then Dj' is inverse transformed using an inverse DCT to obtain dj' = A- ¹ [Dj']:

d'i,k,m,n = 0.25 ∑∑ 0<s,t<7 C(s) C(t) D'ι,k.s,t cos[π(2m+1 )s/16] cos[π(2n+1 )t/1 6] ;

The pixels d'i,k,m,n are stored in a reconstructed quantization frame buffer for future iterations.

In the following example, the transcoder uses the motion vector mj, the inter/intra decision and the loop filter decision received from the H.261 encoder without any modifications. The source format (CIF/QCIF) is also not modified. This greatly simplifies the implementation of the transcoder.

The transcoder converts the transcoded sequence Yj' into a bit stream using a noiseless source encoder as in the H.261 encoder.

The bit stream generated by the transcoder is typically decoded by an H.261 decoder, as shown in FIG. 1 1 , numeral 1100. Here, after recovering the sequence Yj' from the received bit stream as usual, the decoder reconstructs the signal ZJ' = [ZJ, ι ,o,o, Zi.6,7,7] = n + dj + dj ¹ , as follows:

(1 ) The received vector Yj' , typically output by a variable- length decoder (1102), is first transformed using an inverse DCT unit (1 104) that is operably coupled to the variable-length

decoder (1 102), wherein the inverse DCT unit (1 104) determines an inverse DCT to obtain the vector yj' = [y'i,ι ,o,o y'i,6,7,7]:

y'l.k.m.n = 0.25 ∑∑ 0<s,t<7 C(s)C(t) Y'i.k.s.t cos[π(2m+1 )s/16] cos[π(2n + 1 )t/1 6].

An adder (1 106) is operably coupled to the inverse DCT unit (1 104) and to a loop filter (1 108) to combine yj' and pi', and generate output vector z\ Vector ZJ' is an output of the H.261 decoder and is also utilized as a feedback vector to a frame buffer (1 1 12).

(2) A motion compensation unit (1 1 10) is operably coupled to receive the output of the frame buffer (1 112) and a motion vector mi = (mil , nrij2) f ^rom the encoder that is used to determine a 16x16 window Wj. The pixels in the reconstruction buffer that lie in that window are used to form the vector UJ' = [u'j,ι ,o,o

U 'i,6,7,7] .

A loop filter (1 108) is operably coupled to the motion compensation unit (1 1 10) for providing vector ps ¹ to the adder

(1 1 06). The prediction value pi' = [p'j,ι ,o,o P'i,6,7,7] is obtained from Uj' based on the encoding mode of the current MB:

a. In the intra mode, set pj' = 0.

b. In the predictive (inter) mode:

b1. If the loop filter (1 108) is out, set pi' = UJ'. b2. If the loop filter (1 108) is in, filter the elements of UJ' (block-by-block) using a separable, two- dimensional 3-tap FIR filter, and set the output of the filter equal to pi'.

(3) The vector pj' is added to yj' to obtain the reconstructed vector ZJ' = yj' + pi'.

The overall quantization error is then the sum of the quantization errors dj and dj' introduced by the encoder and the transcoder, respectively.

FIG. 12, numeral 1200, is a flow chart of one embodiment of the steps in accordance with the method of the present invention. The method provides for transcoding a sequence of quantized vectors Yj generated by a predictive waveform encoder utilizing modified quantization error vectors Sj. The vectors cited below are described with greater particularity above.

The method comprises the steps of: (1 ) generating a difference vector Ej' = Yj - Sj between at least a first quantized vector Yj and a modified quantization error vector Sj (1202); (2) quantizing the difference vector Ej' to obtain a transcoded vector Yj ¹ (1204); (3) receiving at least two of the vectors Yj, Ej' and Yj' and generating the inverse-transformed error vector dj' = Ar ¹ [Dj'], where Dj' is a quantization error vector and Aj- ¹ is an inverse transformation (1206); and generating the modified quantization error vector Si based on past values of the vector dj' (1208).

As described for the device of the present invention, the transformation Aj- ¹ may be selected to be constant, the quantized vector Yj may have dimensionality L = 1 , and the inverse transformation Ar ¹ [ ] may be an identity matrix.

The step of generating the modified quantization error vector

Sj based on past values of the vector dj' may be selected to include: (1 ) generating a predicted quantization error vector Sj, and (2) utilizing SJ to generate a modified quantization error vector Sj such that Sj = AJ[SJ] where Aj is a predetermined transformation. In addition, this step may include one of (1 )-(2):

(1 ) utilizing received vectors Yj' and Ej', for determining a quantization error Dj' = Yj' - Ej', and utilizing an inverse transformation Ar ¹ to generate the vector dj' according to dj' = Ar ¹ [Dj'], and (2) receiving and utilizing vectors Yj and Yj ¹ for determining an error Xj' = Yj' - Yj, utilizing inverse transformation Aj- ¹ for generating a vector Xj according to Xj = Aj- ^"1 [Xj'], and adding xi and Sj to obtain the vector dj'.

The quantized vector Yj may be generated by a predictive digital video encoder and represents quantized transform coefficients of a macro block (MB), a fundamental coding element in H.261. A motion-compensated prediction (MCP) may be utilized for prediction. Transformation A[ ] is a Discrete Cosine Transform (DCT).

In one embodiment, the video encoder may be selected to be an H.261 encoder.

The present invention may be embodied in other specific forms without departing from its spirit or essential characteristics. The described embodiments are to be considered in all respects only as illustrative and not restrictive. The scope of the invention is, therefore, indicated by the appended claims rather than by the foregoing description. All changes which come within the meaning and range of equivalency of the claims are to be embraced within their scope.

We claim: