DE19509373C1 - Object based coding of image sequences - Google Patents

Abstract

The object based image coding system has an image analyser (1) for object coding and receives three input parameters that are a motion map (2), an original image (3) and a disparity map (4). Images (3,5) obtained by two different cameras are received by a disparity circuit (14). The analyser segments the object of the original image and determines object parameters based upon a source model. The parameters are coded (7) and the values received by a decoder (8) provides colour (C'), a motion vector (B), shape (F) and a disparity vector (D). The values are stored in memory and used in the coding cycle as well as by an image synthesiser (10) to generate an image for comparison (12) with the original. The result is used by a model error coding stage.

Description

A technique that has some disadvantages of block-based Coding of image data, e.g. B. the so-called mosquito effect, Avoids, the object-based coding of image data. Another advantage of this technique is that, in contrast to the block-based coding method no longer the whole Image information of one pixel must be transmitted, but only the parameters of the segmented objects an image. This is through the object-based approach further compression of the image data is achieved. With help object-based coding of video sequences also the content-related processing of the transmitted Allows images in multimedia applications. Possible Object parameters that describe an object of an image are z. B. the position of the object within the image, movement of the object, shape of the object or color of the object.

The main problem with object-based image coding is in the correct segmentation of the image into objects.

Due to the advantages of object-based described above th approach are possible, such as. Tie Areas of video telephony or in various new ones Multimedia applications. In the area of new multimedia Z is used. B. the construction of an interactive image database ken possible.

The process of a first original picture and a second Original image taken by two cameras that a scene from different at the same time Take perspectives, to receive depth information, is called binocular stereo (cf. D. Geiger, B. Landendorf, A. Yuille, Occlusions and Binocular Stereo,  Computer Vision ECC V-92, 2nd European Conference on Computer Vision, Santa Margharita Ligure, Italy, 18-23 May 1992, P.425-433).

In a stereo pair of pictures, a first original picture and a second original image taken by two cameras were, it may be due to the different properties of the cameras and their settings to a general one Imbalance in the brightness values between the first and second channels come. Statistical model This difference and compensation of this sub differences can improve the disparity estimate (see M. Ziegler, W. Tengler, T. Tabeling, Influence of Camera Cali bration on the Coding of Stereo Sequences, Proc. of the 1st International Festival on 3D Images, Paris, France, Sept. 1991, pp. 184-190).

A process for the so-called matching of images using the Disparity information of the pictures is known (S. Barnard and W. Thompson, Disparity Analysis of Images, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. PAMI-2, No. 4, pp. 333-340, July 1980).

Methods for object-based image coding are known which require prior knowledge of the content of the image (ND Duffy. JFS Yan, A Knowledge Based Approach to Image Coding, Signal Processing IV: Theories + Applications, Elsevier Science Publishers BV, Eurasip 1988, p.1125- 1128). Methods that segment the objects based on their movement are also known (M. Hötter "Object-Oriented Analysis-Synthesis Coding Based on Moving Two-Dimensional Objects", Signal Processing: Image Communication Journal Vol. 2, 1990, Elsevier Publications , Pp. 409-428), (J. Ostermann "Object-Based Analysis-Synthesis Coding Based on the Source Model of Moving Rigid 3 D Objects", Image Communication Journal, Vol. 6, 1994, pp. 143-161). These methods can only be used to a very limited extent, e.g. B. in the so-called head and shoulder scenes in videophone sequences.

The required prior knowledge about the content of the pictures and the Fact that in object-based segmentation due to the movement of the objects static image areas, e.g. B. a static background, not in different objects can be segmented, face significant disadvantages method.

The problem underlying the invention is an object-based one Coding of image sequences without prior knowledge of the content of the pictures.

This problem is solved by the method according to claim 1 and the arrangement according to claim 11 solved. With Advantages achieved by the invention are essentially therein to see that object-based processing of Video sequences without prior knowledge of the content of the images is made possible and thus the method can be used in general is without the limitations known procedures subject to. By using a second camera and the resulting stereo information, also disparity called, can also static objects such. B. the image background, segment into objects, what with known method is not possible. By in the dependent steps described for implementation The amount of data to be transferred becomes a control loop significantly reduced since only the changes to the Object parameters are transferred and not all  Object parameters. Also the quality of the object segmentation is improved because the image analysis unit based on the recovered images errors or inaccuracies recognizes and the segmentation of the following images into objects is adjusted accordingly.

Restricting the search area becomes a significant one Computing time savings when performing the dynamic Programming algorithm reached.

Further developments of the invention result from the dependent Claims.

An embodiment of the invention is in the drawings shown and is described in more detail below.

Show it

Fig. 1 is a block diagram showing the entire estimation process for the object-based image processing, the movement, the disparity, the object image analysis with segmentation, the parameter coding and the control loop to improve the transferred image information represents;

Fig. 2 is a block diagram illustrating the entire procedure for object-based image processing, if the receiver is to obtain the possibility of even two images, the reconstructed first image and the second image restored to rekonstruie reindeer;

Fig. 3 is a diagram illustrating an example of an optimal path, which has been using the Dynamic Pro Gramming algorithm three-dimensional Suchbe rich found;

Fig. 4 is two diagrams illustrating how according to claim 13 is moved with the occlusion areas;

Fig. 5 is a flow chart describing the segmentation of the objects;

Fig. 6 is a sketch that describes the process of polygon approximation.

Procedural the invention is ren and further illustrates the arrangement of the invention with reference to FIGS. 1 to 6.

Fig. 1, the entire process for the object-based image encoding describes. An image analysis unit 1 uses the following three input variables for object segmentation: a motion map 2 , the first original image 3 and a disparity map 4 . The first original image 3 is the image that is to be divided into objects.

The image analysis unit 1 segments the objects of the first original image and determines object parameters, each of which describes the segmented objects, according to a known source model 6 . Object parameters are e.g. B. the color of the object C, a motion vector of the object, the shape of the object F and the disparity vector.

The object parameters are made available to a unit for parameter coding 7 , which encodes the object parameters. A parameter decoding unit 8 decodes the encoded object parameters locally, e.g. B. to the decoded object parameters color of the restored object C ', motion vector of the restored object -', shape of the restored object F 'and disparity vector of the restored object'.

The restored object parameters are stored in a memory 9 . From this memory, the restored object parameters are fed to the parameter coding unit 7 so that the restored object parameters can be compared with the object parameters of the following image, in order to only have to code and transmit the change information of the individual object parameters of the following images and thus to save transmission capacities. In addition, the restored object parameters are fed to an image synthesis unit 10 , which reconstructs the first restored image 11 from the restored object parameters.

The restored image 11 is fed to the image analysis unit 1 . A comparison of the restored image with an image immediately following the first original image is carried out in the image analysis unit 1 . Based on the comparison z. B. Changes to the objects discovered and the object segmentation adapted to these changes. Furthermore, the restored image 11 is made available to a comparison unit 12 , which compares the first original image 3 with the restored image 11 and detects any model errors that occurred during the coding.

The model errors discovered by the comparison unit 12 are coded by a model error coding unit 13 and then transmitted to a receiver.

Disparity Estimation and Postprocessing 14

The process of obtaining a first original image 3 and a second original image 5 , which are recorded by two cameras that record a scene at the same time from different perspectives, for obtaining depth information, is called binocular stereo (cf. D. Geiger, B. Landendorf , A. Yuille, Occlusions and Binocular Stereo, Computer Vision ECC V-92, 2nd European Conference on Computer Vision, Santa Margharita Ligure, Italy, 18-23 May 1992, p.425-433).

The disparity estimation algorithm 14 consists of four parts:

• 1. Global modeling and compensation of the differences between the two cameras used, which take the first original image 3 and the second original image 5 .
• 2. Determination of a usable search area for Dispa rity estimate along a scan line.
• 3. Disparity Estimation Based on Dynamic Programming Algorithm based on three-dimensional search area.
• 4. Post-processing of the disparity card 4 .

Global modeling and compensation of differences between between the two cameras used

In a pair of stereo images, a first original image 3 and a second original image 5 , which were taken by two cameras, there can be a general imbalance in the brightness values between the first and the second due to the different properties of the cameras and their settings Channel come. Statistical modeling of these differences and compensation for these differences can improve the disparity estimate (see M. Ziegler, W. Tengler, T. Tabeling, Influence of Camera Calibration on the Coding of Stereo Sequences, Proc. Of the 1st International Festival on 3D Images, Paris, France, Sept. 1991, pp. 184-190). However, this compensation is only global and the quality of the result is limited for two reasons. The two different perspectives from which the first original image 3 and the second original image 5 are recorded by the two cameras lead to different reflections within the images. These different reflections are not taken into account by the model described above. In addition, there are areas on the edges of the first original image 3 and the second original image 5 that do not appear in the other image. Nevertheless, they flow into the statistical model cited. To compensate for these disadvantages, it is necessary to introduce a value that describes the difference in the local brightness value between the two cameras used. The value is taken into account in the cost function of the dynamic programming algorithm. The value is calculated using the following known formula:

The individual elements of the formula have the following meaning:
n, m: pixel coordinates
b (n, m): local compensation factor for large differences in brightness between the two images
N = (τ + 1) _* 3: number of pixels in the edited window
2τ + 1: Width of the processed window in the x direction
W _L (i, j): brightness value of the first original image at the point (i, j)
W _R (i, j): brightness value of the second original image at the point (i, j) (): displacement vector for the pixel with the pixel coordinate n.

Determination of a usable search area for disparity estimation along a scan line

The three axes of the three-dimensional search area of the dynamic programming algorithm (see FIG. 3) are the x axis of the image 15 , the displacement in the x direction d ^x 16 and the displacement in the y direction d ^y 17 . The larger the ranges of the shift values, the larger the search area will be. As a result, more paths have to be taken into account in the calculation for the dynamic programming algorithm. Since the dynamic programming algorithm is very computationally complex, it is advantageous to reduce the number of paths to be taken into account by precalculating the area in which the displacement vectors can be located. This is done by first using a block matching method to estimate the range of values in which the disparity vectors are likely to be found. This reduces the search area for the dynamic programming algorithm. In order to smooth the resulting disparity card 4 , block-based median filtering is applied to the disparity card 4 in order to correlate the individual blocks of the disparity card with the respective neighboring blocks. A block-by-block calculation of the minimum search area and the maximum search area taking into account the neighboring blocks follows the median filtering.

Disparity estimation using Dynamic Programming Algo rhythm with three-dimensional search area

The dynamic programming algorithm uses a so-called cost function, which in this connection evaluates the difference between areas on the two examined images, the first original image 3 and the second original image 5 . The entire cost function is made up of the difference in the brightness values of two examined windows, called the feature difference (FD), and the so-called disparity jump costs.

Feature difference

The windows used to calculate the FD can be found under be of different sizes, it should be noted that a very small window a very small area under investigation represents, the use of which can lead to too many Patterns of a first window can be found in the search area can and therefore supply an incorrect displacement vector. In a window that is too large, there are usually hidden areas. Masking is understood in stereo image processing an area that can only be seen in the first original image is, but not in the second. Make these concealments is a problem with so-called matching. A good compromise in  the window size is e.g. B. a window that is nine pixels wide and is three pixels high. The FD is normalized to one Value between 0 and 1. The extreme value FD = 0 means that a match of the brightness values in the two examined windows exists, and FD = 1 shows a total Inconsistency of the brightness values. The FD will calculated in the following way:

Except for a newly introduced constant c, the one Normalization constant with an empirical value 1/12 describes the terms in this formula in used the same meaning as explained above.

Disparity jump costs

It has proven to be cheap, in addition to the FD in the a regularization factor for the entire cost function take into account the large disparity differences along punished in the horizontal direction. A suitable function is given in (D. Geiger, B. Landendorf, A. Yuille, Occlu sions and Binocular Stereo, Computer Vision ECC V-92, 2nd European Conference on Computer Vision, Santa Margharita Ligure, Italy, 18-23 May 1992, pp.425-433):

Experiments have shown that the values µ = 0.3 and ε = 0.15 deliver the best results.

Entire cost function

The entire cost function results from the sum of the node costs and the path costs. The node costs are the FD _n ().

The path costs result from the following rule:

each represent the absolute disparity vector. The total costs of a node result from:

The values of n-1 _{best are} calculated in the following way:

Once the cost of a node is calculated, building on this cost of the previous node, the additional cost of the paths to the end of the line  be calculated. The disparity and the y vector of each Predecessor node is saved to later a known one To enable backtracking. If you look at the last column the lowest value of the total cost used and in a backtracking process of this Path representing cost selected.

This selected path corresponds to the disparity curve of an image line. The procedure described is to be carried out for each image line. After processing all image lines of an image, the result is the disparity card 4 used in the following.

Post-processing of the disparity card 4

An important prerequisite for the subsequent segmentation The object is the post-processing of the disparity card te resulting from the dynamic programming algorithm three-dimensional search area. The meaning of Post-processing is mainly based on the following points:

• - Incorrect disparities due to correlations Neighbor pixels corrected.
• - The disparity card is smoothed.
• - obscurations caused by the disparity, are based on a method described below fills.

For subsequent image coding, it is very useful to use the To reduce the number of objects as possible and the contours to smooth out so less information to describe the Objects is needed.

Post-processing is divided into three steps:

• 1. A median filtering with a 3 × 3 window.
• 2. Interpolation of the concealments.
• 3. Median filtering with 1 × 5 window.

After median filtering with a 3 × 3 window was carried out in order to eliminate individual pixel structures and directly adjust the disparity of the middle pixel to its correlating neighboring pixels, there is the second Step in a process that occludes in interpolate the disparity map.

Interpolation of concealments

FIG. 4 shows the area of an occlusion 18 and the associated course of the luminance gradient 19 of an image area. For each occlusion, the course of the luminance gradient 19 is examined and the local maximum of the luminance gradient 20 is used to fill the image area between the location of the local maximum of the luminance gradient 20 and the adjacent object with the luminance value of the adjacent object. The image shows an example of the filling of an area 21 between the local maximum of the luminance gradient 20 and an area 22 of an adjacent object with the luminance value 144 with the luminance value 144 . Another area 23 , which is located between the local maximum of the luminance gradient 20 and another area 24 with the luminance value 132 , is filled with the luminance value 132 .

This processing is accompanied by median filtering a 1 × 5 window that is used to display single hori to eliminate zonal line structures. These are through the horizontal processing of the disparity arose and must be eliminated again.

Motion estimation 25

In principle, the same method is used for motion estimation 25 as for disparity estimation 14 .

The result of the motion estimation 25 is made available to the image analysis unit 1 in the form of a motion card 2 , corresponding to the disparity card 4 .

Image analysis

A 2 1/2-dimensional area with fol defining properties:

• 1. The areas have the same motion vectors 25 .
• 2. The areas have the same disparity vectors 14 .

These two properties are used to define one Used as a very good prerequisite for form the further segmentation. These properties will also to determine the relationship between consecutive images needed. In case of a Stereo encoding, which will be described later, are these Properties the basis to switch between a first view and a second view.

Segmentation of the objects

The segmentation of the objects is divided into five steps (see Fig. 5):

• 1. Segmentation of objects 26 .
• 2. Working out the object properties 27 .
• 3. Elimination of small objects 28 .
• 4. Fusion of objects 29 .
• 5. Determination of the object parameters for describing the objects 30 .

A histogram classifier is used to work out objects that have the same disparity vector and the same motion vector. The elements that are used for segmentation are combined in a vector field and denote in detail the disparity d _{i, j} , the horizontal movement h _{i, j} and the movement in the vertical direction v _{i, j} .

The segmentation of the objects is controlled by a mask The areas of the image that are used for segmentation Question come, definitely. The areas on this mask marked that represent model errors and areas that have remained uncovered by themselves from this area moving objects from previous pictures. In the first Original mask marks the entire image as segmentable.

Disparity is the main differentiator. The disparity gives a measure that is proportional to the Depth is and with which it is possible to reconstruct the Images in hierarchical order, starting with the Background objects to build.

Small Object Elimination 28

To reduce the number of segmented objects eliminated small objects by following one below described criterion assigned to the optimal neighboring property be net. In this context, means a small one Object an object whose maximum size is in the range of approx. 100 pixels. The following two conditions must be met be satisfied so that an object is a larger neighboring one Object is assigned:

• 1. (S _j ≠ ERASED, FAULTY) ˆ (A _j X) = TRUE
• 2. (S _j = S _n ) ˆ (A _j A _n ) = TRUE

With X in this context is the size of an object referred to, up to which an object is a small object is looked at.

For each object to be eliminated after the two before qualifications, the following will occur  Criterion used to find out the neighboring object the object to be eliminated should be assigned:

The elements of the previous formulas are as follows explains:

The neighboring object in which the above-described criterion Q _{j, n} becomes minimal is assigned the object to be eliminated. The process of elimination proceeds iteratively and continues until no object fulfills the requirements for elimination.

Fusion of objects 29

To further reduce the number of objects, be Objects merged together, the following condition fulfill:

(S₁ ≠ ERASED, FAULTY) ˆ (S₂ ≠ ERASED, FAULTY) ˆ
(S₁ = S₂) ˆ (D₂ = D₁ ± 1) ˆ (H₁ = H₂) ˆ (V₁ = V₂) = TRUE

Should two or more neighboring objects meet this criterion another criterion is used:

The object that reaches a maximum value for Q is that best object and will be merged. Merging is also an iterative process and is repeated until none Object more meets the requirements for the merger.

The two steps of eliminating small objects 28 and merging objects 29 can reduce the number of objects by a factor of 10 and more, which results in a large saving in computing time and the required object parameters that have to be transmitted.

Description of the objects 30

The contours of an object are first subjected to an edge detection process. Four vertices 31 , 32 , 33 and 34 are extracted from the resulting contour. Starting from the four vertices 31 , 32 , 33 , and 34 , an iterative process for approximating a polygon is started, which determines the remaining vertices of the resulting polygon of the object. Fig. 6 describes this known procedure.

Two vertices are connected by a straight line. A separation value between the point and the line is calculated for each point within the contour between two vertices. If this value lies above a freely selectable threshold, the point is defined as a new vertex 35 for the polygon approximation. This process is repeated until no more vertices are inserted. After completing this process, a list of the previously calculated vertices represents the polygon approximation.

Due to the polygon approximation it can happen that Objects overlap points. The problem can be solved by the stereocopic information can be solved as the objects can be reconstructed according to their depth.

Parameter coding 7

Instead of information about individual pixels or blocks, such as in the case of block-oriented coding methods, the object based coding different parameters for each object, that characterize the object. These parameters are z. B. the color of the object C, the movement of the object B, the shape of the object F and the disparity of the object D. These parameters are encoded and transmitted to the receiver, which then returns the individual objects from the parameters can reconstruct. By reconstructing the objects the receiver also restore the actual image.

Control loop to improve object-based Image encoding

So that any errors occurred in the segmentation and Coding of objects easier to find and correct a control loop is introduced, which consists of the following Components are:

• A unit for decoding the object parameters 8 ;
• - A unit for storing the decoded object parameters 9 ;
• - A unit for the reconstruction of the image from the decoded, stored object parameters 10 ;
• A unit that compares the first original image with the first restored image and finds possible model errors 12 ;
• - A unit for coding the model errors that are transmitted to the receiver 13 .

The coder contains a unit that decodes the transmitted object parameters locally in the coder in order to simulate the receiver in the coder itself. The decoded objects are stored in the encoder. The restored parameters are made available to the parameter coding unit 7 so that, in the case of a sequence of images, only the change in the object parameters of objects of two images that follow one another in time has to be transmitted. The restored parameters are also reconstructed in an image synthesis unit 1 to a restored image 11 . The image synthesis from the objects takes place hierarchically, taking disparity into account. B. Objects that overlap due to movements, are reconstructed in the correct order, so that the object, which was spatially closer to the camera when recording, is also brought back to the foreground during the reconstruction. The restored image is made available to the image analysis unit which, for. B. changes in the objects due to movement and disparity changes are taken into account, and based on this information new objects are segmented or an existing object segmentation is updated. Furthermore, the first restored image 11 is compared with the first original image 3 in order to discover model errors. These areas of the model errors are fed to a unit for model error coding 13 and are treated there as independent objects.

Coding stereo images (see Fig. 2)

In the method described so far, information was transmitted which only enables the receiver to reconstruct the first restored image 11 . A reconstruction of the second restored image 36 is not possible on the basis of the transmitted information, since e.g. B. the information about the masking areas is lost (see section interpolation of the masking). If, however, as described in FIG. 2, the second restored image is compared with the second original image in a further unit for extracting model errors 37 , the entire masking information can be obtained in the unit for model error coding 13 and thus made available to the receiver . So that the receiver can also reconstruct the first restored image and the second restored image 36 separately, it is not necessary to transmit additional information in addition to the additional model errors of the second original image in comparison to the coding of only one image. This enables the recipient to reconstruct two images separately and thus achieve three-dimensional effects. This advantage means that little additional information is transmitted. This saves a lot of transmission capacity.

Claims (13)

1. method for object-based image coding,

• - a motion map ( 2 ) for the first original image ( 3 ) is generated by a motion estimation ( 25 ) of two temporally successive images, a first original image ( 3 ) and a subsequent original image ( 37 ),
• - With a disparity estimate ( 14 ) of the first original image ( 3 ) and a second original image ( 5 ) of the same scene at the same time, taken from different perspectives, a disparity map ( 4 ) for the first original image ( 3 ) and the second original image ( 5 ) is generated,
• - in which an image analysis unit ( 1 ) from the first original image ( 3 ), the movement map ( 2 ) and the disparity map ( 4 ) segments image areas of the same motion estimate and the same disparity estimate into objects of the first original image,
• in which the image analysis unit ( 1 ) determines object parameters (C, F and) for the objects that characterize the objects,
• - in which the object parameters (C,, F and) are coded in a unit for coding the object parameters ( 7 ),
• - In which the coded object parameters ( 38 ) are transmitted.

2. The method according to claim 1, in which a regulation for improving the transmitted object parameters is provided, the following steps being provided for the regulation:

• the coded object parameters are decoded in a unit for parameter decoding ( 8 ),
• - the decoded object parameters (C ′, ′, F ′ and ′) are saved ( 9 ),
• - The decoded object parameters (C ',', F 'and') are compared with the object parameters of the objects of a temporally following the first original image- first original image in the parameter coding unit ( 7 ) and only the changes in the object parameters between the first original image ( 3 ) and the subsequent first original image are transmitted.

3. The method according to claim 2, in which the following additional steps are provided for the control:

• a first restored image ( 11 ) is reconstructed from the decoded object parameters (C ′, ′, F ′ and ′) in an image synthesis unit ( 10 ),
• - The first restored image ( 11 ) is made available to the image analysis unit ( 1 ),
• - in the image analysis unit ( 1 ) the first restored image ( 11 ) is compared with the first original image ( 3 ),
• changes in the objects due to movement or disparity are analyzed in the image analysis unit ( 1 ),
• - When segmenting the objects for a subsequent image, the changes are taken into account in order to improve the segmentation of the objects.

4. The method according to claim 3, in which the following additional steps are provided for the control:

• a first unit for discovering model errors ( 12 ) calculates model errors ( 39 ) from a comparison of the first original image ( 3 ) with the first restored image ( 11 ),
• - The model errors ( 39 ) are encoded in a unit for model error coding ( 13 ).

5. The method according to any one of claims 1 to 4, wherein the disparity estimate ( 14 ) in the form of a dynamic programming algorithm with a three-dimensional search area ( 15 , 16 , and 17 ). 6. The method according to claim 5, wherein the disparity estimate ( 14 ) provides steps for preprocessing image data which reduce the search area ( 15 , 16 , and 17 ) for the dynamic programming algorithm. 7. The method according to any one of claims 1 to 6, in which steps are provided in the segmentation of the objects for eliminating a first object ( 28 ), the first object being assigned to a larger neighboring second object according to the criterion Q _{j, n} denotes the criterion according to which an object j adjoins an adjacent object n when n = 1. . in neighboring objects, where this criterion results in a minimum value,
D _{j describes} the disparity vector of object j and D _{n describes} the disparity vector of object n,
H _{j describe} the horizontal movement vector of the object j and H _n the horizontal movement vector of the object n,
V _j the vertical motion vector of the object j and V _n the vertical motion vector of the object refer to n, and
B _{j, n} indicates the number of pixels that make up a common contour of the object j and the object n. 8. The method according to any one of claims 1 to 7, wherein the segmentation of the objects steps are provided for merging ( 29 ) adjacent objects, which determines the objects that are to be merged according to the following criterion: (S _j ≠ ERASED, FAULTY) ˆ (s _j ≠ ERASED, FAULTY) ˆ (S _i = S _j )
(D _j = D _i ± 1) ˆ (H _i = H _j ) ˆ (V _i = V _j ) = TRUE, where S _{i describe} the status of the object i and S _j the status of the object j with that used in this context ERASED, which indicates that the object is merged with another object, and FAULTY, which indicates that the object cannot be restored.
D _{j describes} the disparity vector of object j and D _{n describes} the disparity vector of object n,
H _{j describe} the horizontal motion vector of the object j and H _n the horizontal motion vector of the object n, and
V _j denotes the vertical movement vector of the object j and V _n the vertical movement vector of the object n. 9. The method according to any one of claims 3 to 8, in which the following additional steps are provided for the control:

• - The first restored image ( 11 ) and a second restored image ( 36 ) are reconstructed in the image synthesis unit ( 10 ),
• a second unit for discovering model errors ( 40 ) calculates model errors ( 41 ) of the second original image ( 5 ) from a comparison of the second original image ( 5 ) with the second restored original image ( 36 ),
• - The model errors of the first restored image ( 39 ) and the model errors of the second restored image ( 41 ) are encoded and transmitted.

10. The method according to any one of claims 1 to 8, in which the following additional steps are provided in the post-processing of the disparity estimate ( 14 ):

• - For areas of the disparity map ( 4 ) in which occlusion ( 18 ) occurs, the local maximum of the luminance gradient ( 20 ) is searched for,
• - In each case a partial area of the masking ( 18 ) from the local maximum of the luminance gradient ( 20 ) to an adjacent object is filled with the luminance value ( 21 and 23 ) of the object adjacent to the partial area ( 22 and 24 ).

11. Arrangement for performing the method according to one of claims 1 to 8 or 10,

• a unit for estimating the movement of the first original image ( 25 ) on the basis of the first original image ( 3 ) and the following original image ( 37 ) is provided,
• a unit for estimating the disparity of the first original image ( 14 ) on the basis of the first original image ( 3 ) and the second original image ( 5 ) is provided,
• - In which a unit for image analysis, object segmentation and for determining the object parameters of the first original image ( 1 ) is provided, and
• in which a unit for object parameter coding ( 7 ) of the objects segmented by the arrangement for image analysis ( 1 ) is provided,

12. Arrangement according to claim 11,

• a unit for estimating the movement of the first original image ( 25 ) on the basis of the first original image ( 3 ) and the following original image ( 37 ) is provided,
• in which a unit for decoding the object parameters of the object parameters ( 8 ) is provided for the reconstruction of the first original image ( 3 ),
• - in which a unit for storing the object parameters ( 9 ) is provided,
• - in which a unit for image synthesis ( 10 ) of the first original image ( 3 ) from the decoded object parameters of the objects of the first original image (C ',', F 'and') is provided,
• - In which a unit ( 12 ) for comparing the first original image ( 3 ) with the restored first image ( 11 ) and for recognizing the model errors is provided, and
• - in which a unit for coding the model errors is provided ( 13 ).

13. Arrangement for performing the method according to claim 9,

• a unit for estimating the disparity of the first original image ( 14 ) on the basis of the first original image ( 3 ) and the second original image ( 5 ) is provided,
• - in which a unit ( 1 ) for image analysis, object segmentation and for determining the object parameters of the first original image is provided,
• a unit ( 7 ) for coding the object parameters of the objects (C, F and) segmented by the arrangement for image analysis ( 1 ) is provided,
• - in which a unit ( 8 ) for object parameter decoding of the object parameters for the reconstruction of the first original image ( 3 ) is provided,
• - in which a unit ( 9 ) is provided for storing the object parameters,
• - in which a unit ( 10 ) for image synthesis of the first original image from the decoded object parameters (C ',', F 'and') of the objects of the first original image is provided,
• a unit ( 12 ) for comparing the first original image ( 3 ) with the restored first image ( 11 ) and for recognizing the model errors ( 39 ) is provided,
• - In which a unit ( 13 ) is provided for coding the model errors, and
• - A unit ( 40 ) for comparing the second original image ( 5 ) with the restored second original image ( 36 ) and for recognizing the model errors ( 41 ) is provided.