VIDEO TRANSCODING WITH REDUCED DRIFT

1. Scope of the invention

The invention relates to a transcoding device of a first bitstream (S1 ) into a second bitstream (S2), said bitstreams being representative of a same sequence of pictures. More particularly, the invention relates to a transcoding device of the FPDT type (Fast Pixel Domain Transcoder).

2. Prior art A transcoding device is used to modify the coding cost of a sequence of pictures. Indeed, it is sometimes necessary to transfer a bitstream representative of a sequence of pictures of a first network of bandwidth B1 to a second network of bandwidth B2, where B1 > B2. For this purpose, a transcoding device is used to modify the coding cost of said sequence of pictures, i.e. the number of bits used to encode it. Such a transcoding device also enables a bitstream to be adapted to the resources of a terminal or even such a bitstream to be inserted into a multiplex.

A transcoding device 1 of the FPDT type according to the prior art is shown in figure 1. It is notably described by G. J. Keesman, in the document entitled "Multi-program Video Data Compression", Thesis Technische Universitat Delft. ISBN 90-74445-20-9, 1995. Such a transcoding device receives at its input a first bitstream S1 representing a sequence of pictures. The input of the transcoding device is connected to an entropy decoding module VLD, itself connected to a first inverse quantization module IQ1. The decoding module VLD decodes part of the first bitstream into current picture data I that are then dequantized by the first dequantization module IQ1 into dequantized data ID with a first quantization step. This first quantization step is itself decoded from the bitstream S1. In general, the picture data I is in the form of coefficient blocks. The quantization module IQ1 is connected to a first input of a first computation module C1. The first computation module C1 is suitable to calculate residual data R. For this purpose, the first computation module C1 carries out the difference between the current dequantized data ID and prediction data PT sent to a second input of the first computation module C1. The output of the first computation module C1 is connected to the input of

a quantization module Q2 suitable to quantize the residual data R into quantized residual data RQ with a second quantization step. The second quantization step is determined according to the required bit rate B2. This quantized residual data RQ is then transmitted to an entropy coding module VLC to generate part of the second bitstream. It is also sent to a second dequantization IQ2 operating an inverse quantization from the one operated by the quantization module Q2 and generating dequantized residual data RD. This dequantized residual data RD is then transmitted to a first input of a second computation module C2. The second computation module C2 is suitable to compute requantization errors data E. For this purpose, the second computation module C2 carries out the difference between the dequantized residual data RD and the corresponding residual data R sent to a second input of the second computation module C2. The output of the second computation module C2 is connected to the input of a first IDCT transformation module applying a first transform to the requantization errors data E to generate requantization errors in the spatial domain also called pixel domain, called transform requantization errors data EP. The IDCT module preferentially operates an Inverse Discrete Cosine Transform. The output of the first IDCT transformation module is connected to a memory MEM. Generally, this memory MEM stores over 8-bits each of the data at the output of the first transformation module. The memory is also connected to a prediction module PRED suitable to generate intermediate prediction data P from the transform requantization errors data EP stored in the memory MEM. The prediction module PRED implements, for example, a temporal prediction by motion compensation using motion vectors MVs decoded from the bitstream S1 in the case where the current dequantized data ID is in INTER mode. It can also implement a spatial prediction for example in the case where the current dequantized data is data in INTRA mode as defined in the video coding standard H.264. The intermediate prediction data P is then sent to the input of a second DCT transformation module that applies a second transform to said intermediate prediction data P to generate the prediction data PT. The DCT module preferentially operates a Discrete Cosine Transform. Such a transcoding device 1 has the disadvantage of leading to an effect of temporal or spatial drift. Indeed, the estimation of requantization

errors made while transcoding picture data that serve as temporal or spatial reference to other picture data is not perfect. A drift is introduced that accumulates along a group of pictures known under the acronym GOP leading to a gradual degradation in the quality of said pictures. Figure 2 shows a graph in which curves are shown characterizing the performances obtained with an architecture according to the prior art. The abscissa shows the number of decoded pictures of the sequence and the ordinate the signal to noise ratio or PSNR in decibels of the said decoded pictures with respect to the source pictures. The reference curve 20 corresponds to the decoded pictures of the bitstream S1. The curve referenced 21 corresponds to the decoded pictures of the bitstream S2 at the output of the transcoding device 1 according to the prior art. The drift effect is observed on the curve 21. The quality of the pictures gradually deteriorates until the transcoding of an INTRA type picture.

3. Summary of the invention

The purpose of the invention is to compensate for at least one disadvantage of the prior art. More particularly, the purpose of the invention is to reduce the drift effect of the device according to the prior art. The invention relates to a transcoding device of a first bitstream into a second bitstream, the two bitstreams being representative of a same sequence of pictures. The device comprises:

- means for decoding the first bitstream into picture data,

- means for dequantizing, with a first quantization step, the picture data into dequantized data,

- means for calculating residual data by subtracting prediction data from the dequantized data,

- means for quantizing the residual data with a second quantization step into quantized residual data, - means for dequantizing the quantized residual data into dequantized residual data, and

- means for coding the residual quantized data.

The device also comprises, to calculate the prediction data:

- means for calculating the requantization errors data by subtracting the residual data from the dequantized residual data,

- first transformation means for transforming the requantization errors data into transform requantization errors data, - storage means for storing the transform requantization errors data,

- means for calculating intermediate prediction data from the stored transform requantization errors data,

- second transformation means for transforming the intermediate prediction data into the prediction data. According to a particularly advantageous characteristics of the invention, the first transformation means are suitable to transform the requantization errors data into transform requantization errors data having a resolution strictly less than 1.

According to another particular characteristic, the second transformation means are suitable to transform the intermediate prediction data having a resolution strictly less than 1.

According to another particular characteristic, the memory is suited to store the transform requantization errors data over 8 bits or more than 8 bits

with a resolution equal to — with n>0.

T Advantageously, the first transformation means are suited to transform the requantization errors data in such a manner that a transform requantization errors data of value v1 is stored in the storage means by a value v2, v2 being linked to v1 by the following equation: v1 = (v2-128)/2 ⁿ .

4. List of figures

The invention will be better understood and illustrated by means of advantageous embodiments and implementations, by no means limiting, with reference to the figures attached in the appendix, wherein: figure 1 illustrates a transcoding device according to the prior art, - figure 2 shows a graph on which are drawn the PSNR curves associated respectively with a sequence of pictures decoded from a bitstream S1 and with the same sequence after transcoding of the bitstream S1 by a device of the prior art,

figure 3 illustrates a transcoding device according to the invention, and figure 4 shows a graph on which are drawn the PSNR curves associated respectively with a sequence of pictures decoded from a bitstream S1 , with the same sequence transcoded by a device according to the prior art and with the same sequence transcoded by a device according to the invention.

5. Detailed description of the invention

With reference to figure 3, the invention relates to a transcoding device 2. The transcoding device according to the invention is suited to transcode a first bitstream S1 at a first bitrate B1 into a second bitstream S2 at a second bitrate B2, said bitstreams being representative of a sequence of pictures. The digital pictures are tables of image points or pixels coded on 8 bits. Three components are generally coded: one luminance component and two chrominance components. Each component can therefore take 255 different values.

The modules of the transcoding device 2 identical to those of the transcoding device 1 are identified in figure 3 using the same references and are not described further. The transcoding device 2 comprises the modules VLD, IQ1 , C1 , Q2, IQ2, C2, PRED and VLC. The transcoding device 2 also comprises a first transformation module IDCT2 as a replacement for the module IDCT1. The first transformation module IDCT2 is suited to apply a first transform on the requantization errors data E in such a manner as to generate transform requantization errors data EP in the spatial domain with a resolution (also called precision) strictly less than 1. The resolution is defined as the smallest interval existing between two data, in this case between two transform requantization errors data generated by the first transformation module IDCT2. Preferentially, the first transform is an inverse discrete cosine transform. Classically, such a transform that is implemented using adders, subtracters and shift registers gives rise to rounding errors. For example, the last processing operations of the IDCT function described in the MPEG4 AVC standard are right shifts of 6 bits. By limiting this to a right shift of 4 bits, transform requantization errors data EP is obtained with a resolution equal to

1/4. It is modified so that it generates transform requantization errors data EP

with a resolution equal to — with n>0.

T

The transcoding device 2 also comprises a second transformation module DCT2 as a replacement for the module DCT1. The second transformation module DCT2 is suited to apply a second transform to the prediction data P in such a manner as to generate transform prediction data

PT in the frequency domain from prediction data P having a resolution strictly less than 1. Preferentially, the second transform is a discrete cosine transform. Classically, such a transform that is implemented using shift registers and gives rise to rounding errors. It is modified so as to generate transform prediction data PT from prediction data P having a resolution equal

to — with n>0. According to a particular embodiment, n=2.

According to a first embodiment each of the transform requantization errors data EP is stored in the memory MEM2 over more than 8 bits. According to this embodiment, it is necessary to adapt the prediction module PRED2 so that it can operate on data stored over more than 8 bits, e.g. 10 bits.

According to a particularly advantageous embodiment, each of the transform requantization errors data EP is stored in the memory MEM over unsigned 8 bits with a resolution equal to 1/2 ⁿ . If n=1 , one may stores in MEM errors data with a resolution equal to 14 (e.g. the values -0.5, 0, 0.5, 1.0, 1.5 ....). If n=2, one may stores in MEM errors data with a resolution equal to 1/4 (e.g. values -2.5, -0.25, 2.25, 3.5, 3.75, 4.0, 5.25, 6.5 ....). If n=0, one may only stores in MEM errors data with a resolution equal to 1 (e.g. -4, -1 , 0, 1 , 7 ...). This embodiment is particularly advantageous to the extent that it can use the prediction module PRED of a transcoding device according to the prior art that is suitable to operate on data stored over 8 bits. This embodiment is particularly advantageous in the case where the transform requantization errors data EP has a limited range values, e.g. [-32 ; +31.75]. In this case, if n=2 is chosen, a transform requantization errors data EP of value v1 used by the prediction module PRED is stored in memory MEM2 by a value v2 such

that v1 = (v2-128)/4. Indeed, a value v1 =(2 ⁸ -1 -128)/4=+31.75 corresponds to the value v2=2 ⁸ -1 stored in memory MEM2 and a value v1 =(-128)/4=-32 corresponds to the value v2=0 stored in memory MEM. Hence, the transform requantization errors data EP in the slot [-32 ; +31.75] is stored in memory MEM over unsigned 8 bits with a resolution equal to %.

Figure 4 shows the same graph as that of figure 2, to which the curve 22 has been added. This curve 22 corresponds to the decoded pictures of the bitstream S2 at the output of the transcoding device 2 according to the invention. The signal to noise ratio is improved in the case where a transcoding device 2 according to the invention is used. Moreover, the drift effect characteristic of an FPDT type transcoding is reduced. Indeed, the device according to the invention reduces the rounding errors due to DCT1.