EP0527947A1 - Procede et appareil de decodage par quantification de seuil en treillis a codage par entropie hierarchique pour compression image et video - Google Patents

Procede et appareil de decodage par quantification de seuil en treillis a codage par entropie hierarchique pour compression image et video

Info

Publication number
EP0527947A1
EP0527947A1 EP19910911119 EP91911119A EP0527947A1 EP 0527947 A1 EP0527947 A1 EP 0527947A1 EP 19910911119 EP19910911119 EP 19910911119 EP 91911119 A EP91911119 A EP 91911119A EP 0527947 A1 EP0527947 A1 EP 0527947A1
Authority
EP
European Patent Office
Prior art keywords
image
data
level
encoding
lattice
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Withdrawn
Application number
EP19910911119
Other languages
German (de)
English (en)
Other versions
EP0527947A4 (en
Inventor
Richard L. Baker
Jeffrey Bernstein
Bernd Girod
Xiancheng Yuan
Edmund Thompson
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Polycom Inc
Original Assignee
Picturetel Corp
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Picturetel Corp filed Critical Picturetel Corp
Publication of EP0527947A1 publication Critical patent/EP0527947A1/fr
Publication of EP0527947A4 publication Critical patent/EP0527947A4/en
Withdrawn legal-status Critical Current

Links

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T9/00—Image coding
    • G06T9/008—Vector quantisation
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
    • H04N19/169—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the coding unit, i.e. the structural portion or semantic portion of the video signal being the object or the subject of the adaptive coding
    • H04N19/186—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the coding unit, i.e. the structural portion or semantic portion of the video signal being the object or the subject of the adaptive coding the unit being a colour or a chrominance component
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/50—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using predictive coding
    • H04N19/503—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using predictive coding involving temporal prediction
    • H04N19/51—Motion estimation or motion compensation
    • H04N19/577—Motion compensation with bidirectional frame interpolation, i.e. using B-pictures
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/60—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using transform coding
    • H04N19/61—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using transform coding in combination with predictive coding
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/60—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using transform coding
    • H04N19/63—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using transform coding using sub-band based transform, e.g. wavelets
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/90—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using coding techniques not provided for in groups H04N19/10-H04N19/85, e.g. fractals
    • H04N19/94—Vector quantisation

Definitions

  • the invention relates generally to data communication and signal processing methods and apparatus, and in particular to a method and apparatus for reliably and efficiently encoding and decoding sequences of image data, for example, that transmitted over a telephone communica ions channel.
  • receiver-based motion compensation systems typically are either receiver-based motion compensation systems or transmitter-based motion compensation systems.
  • the receiver makes a prediction as to the motion and compensates the previous frame for the expected motion.
  • the transmitter operating in the same manner, then sends only an error signal describing what must be done at the receiver in order to correct the receiver predicted frame.
  • the error signal is typically coded to reduce its bandwidth.
  • the motion estimation process occurs only at the transmitter. Displacement vectors are generally determined over various regions of the image and this data is then transmitted to the receiver along with an error information data signal.
  • the compensation process is performed on the previously coded image first using the motion information provided by the transmitter.
  • the error signal data provided by the transmitter is then added to the thus compensated receiver image in order to maintain picture quality.
  • each vector is associated with a specific region or block of the image.
  • the blocks are typically non-overlapping and have, for example, a size of eight picture elements (pixels) by eight picture elements.
  • Various methods have been employed for encoding the motion compensation data associated with each of the blocks.
  • Both lattice and vector quantization map a multidimensional space into a finite or countable set of points.
  • vector quantization as described in Ericsson, U.S. Patent 4,849,810, a codebook containing a finite set of points is constructed by some training method or analysis. Each source vector is quantized by locating the "closest" point (or codevector) in the codebook, as measured by some distortion criterion. For example, a squared error criterion (L-2 norm), can be used, although other measures are equally valid.
  • L-2 norm squared error criterion
  • a lattice is defined as a finite set of points and all possible linear translates of it, yielding a (typically) countable but infinite set of points.
  • a source vector can be quantized in accordance with lattice quantization by locating the closest point to it, that is contained in the lattice, as measured by some distortion criterion (for example a squared error criterion) . Assuming the lattice can be indexed (counted) by some method, the lattice point's index is transmitted and the receiver approximates the source vector with that lattice point.
  • Another object of the invention is to control the number of bits employed to transmit each image of a sequence of images while providing for graceful degradation of the image during a scene change or during periods of heavy motion.
  • Other objects of the invention are a motion compensation encoding and decoding method and apparatus which reliably transmit and receive an accurate estimate of the displacement error in a scanned image in a sequence, and an improved motion estimation error encoding and decoding method and apparatus which enable real-time, accurate determination of regional displacement in an image transmission device.
  • the invention relates to a method and apparatus for encoding interframe error data in an image transmission system, and in particular, in a motion compensation image transmission system, for transmitting a sequence of image frames from a transmitter station to a receiver station.
  • the method features the steps of decimating an interframe predicted image data representing a prediction of the current image frame for generating a prediction pyramid data structure representing the current image prediction and having a plurality of decimation levels; decimating an uncoded current image data representing the current uncoded image frame for generating a current image pyramid data structure representing the current image and having the plurality of decimation levels; and applying an entropy coded, lattice threshold quantization encoding method to the difference of the prediction and current image pyramid data structures, on a level by level basis, for generating an encoded data representation of the difference between the predicted image data and the uncoded current image data.
  • the method features the steps of applying the lattice encoding method to the data structures of a level on a block-by-block basis using an E-8 lattice structure.
  • the method features the steps of forming a difference image representing, on a pixel-by-pixel basis, the difference between a predicted image data for a current image frame and an uncoded current image data representing the uncoded current image frame.
  • the method further features decimating the difference image for generating a difference image pyramid data structure having a plurality of decimation levels and selectively applying the lattice quantization encoding to the difference image pyramid data structure on a level-by-level basis for generating an encoded data representation of the difference between the predicted image data and the uncoded current image data.
  • the method features forming the predicted image data using interframe motion compensation.
  • the apparatus of the invention features circuitry for decimating the interframe predicted image data for a current image frame for generating the prediction pyramid data structure having a plurality of decimation levels, circuitry for decimating the uncoded current image data representing the current uncoded image frame for generating a current image pyramid data structure having the plurality of decimation levels, and circuitry for applying entropy coded lattice threshold quantization encoding to the prediction and current image pyramid data structures on a level-by-level basis for generating an encoded data representation of the difference between the predicted image data and the encoded current image data.
  • the apparatus further features circuitry for applying the lattice encoding method to the data structures of a level on a block-by-block basis using an E-8 lattice.
  • the apparatus for encoding the interframe error data in an image transmission system for transmitting a sequence of image frames features circuitry for forming a difference image representing, on a pixel-by-pixel basis, the difference between predicted image data for a current image frame and an uncoded current image data representing an uncoded image frame.
  • Decimation circuitry is provided for decimating the difference image; for generating a difference image pyramid data structure having a plurality of decimation levels; and for applying entropy coded, lattice threshold quantization encoding to the difference image pyramid data structure on a level-by-level basis for generating an encoded data representation of the difference between the predicted image data and the encoded current image data.
  • Figure 1 is an electrical block diagram of a typical image communications system in accordance with the claimed invention
  • Figure 2 is an electrical block diagram of the transmitter of a motion-compensated image encoding apparatus employing the invention
  • Figure 3 is an electrical block diagram of the receiver of a motion-compensated image encoding system for receiving the channel signals from the transmitter of Figure 2;
  • Figure 3A is a block diagram of the lossy compressor 28 according to the invention.
  • FIG. 4 is a block diagram of the lossy compressor 46 according to the invention.
  • Figure 5 is a diagrammatic representation of a one dimensional decimation process
  • FIG. 6 is a detailed electrical block diagram of lossy compressor 46 according to the invention.
  • Figure 7 is a general block diagram of an alternate embodiment of lossy compressor according to one aspect of the invention.
  • Figure 8 is a diagrammatic representation of the relative locations of adjacent blocks used in predicting a next block value
  • Figure 8A is a diagrammatic representation of the relative locations of adjacent motion vectors used in the linear predictor
  • Figure 9 is a probability density function divided into stepped regions and showing centroid locations
  • Figure 10 is an electrical block diagram of a coding apparatus according to the invention.
  • Figure 11 is an electrical block diagram of an entropy coded, lattice threshold quantization encoding system according to the invention. Description of Particular Preferred Embodiments
  • a communications system 6 has a transmitter 8 which, in accordance with a preferred embodiment of the invention, has a camera 10 for providing a video signal to an analog-to-digital converter and frame buffer 12.
  • the frame buffer portion of the analog-to-digital converter and frame buffer 12 is capable of storing a full frame of the video, sampled to, for example, eight bits across a 256x240 pixel raster.
  • the entire coding and motion compensation process takes place in the digital domain.
  • the transmitter has an error signal circuitry 14 and a motion estimation and coding circuitry 16.
  • a channel encoder 18 channel encodes the outputs of the error circuitry 14 and motion estimation and coding circuitry 16 and passes the thus encoded data onto a channel 20 for transmission to a receiver 21.
  • the illustrated motion estimation and coding circuitry 16 of Figure 1 compares a present original input frame image, available over lines 22, with the previous original input image, available in this illustrated embodiment from a frame buffer 24.
  • a motion estimator circuitry 26 generates a measure of the motion displacement between the frames input thereto, and can be any of a number of motion estimation devices as are well known in the art.
  • the motion estimation device uses an adaptive steepest descent error minimization method to generate the motion displacement measures as described in Hinman, U.S. Patent No. 4,661,849.
  • the output of the motion estimator 26 is a field of motion vectors which, as noted above, provide a measure of the motion displacement between input frames.
  • This vector field provides a description of how to map a previous input frame or input image from buffer 24 into the best approximation of the, present input frame or image over lines 22.
  • buffer 24 contains only 32x30 and 64x60 decimated copies of the previous image, while a buffer 44 is used for deriving a 128x120 copy of the previously (estimated) image.
  • best is meant an error metric such as, for example, a mean-squared-error error measure.
  • the motion estimator uses a region matching technique between non-overlapping blocks of the previous and present input images.
  • the estimator will determine which block in the previous image is the best match for the block in the present image, and the value of the displacement is the difference between a new coordinate pair for the block in the present image and the original coordinate pair for the block in the earlier image. That determination defines the motion vector to be associated with the block in the present image.
  • the preferred embodiment of the invention modifies the motion vector field, thereby losing some information, to facilitate the compression of the motion representing data.
  • this operation is represented by a "lossy compressor" 28 which reduces -li ⁇
  • the lossy compressor circuitry 28 is employed for coding the motion vector field available over lines 32, and provides, over lines 30, a coded motion signal representative of the motion vectors.
  • This output of the lossy compressor will not, upon decoding, exactly reproduce the signals over lines 32 (which provide the measure of motion displacement) and, therefore, have some error signal associated with them.
  • the reduction in the data requirements of a lossy compressor when compared to, for example, a PCM exact coding method, are so substantial, that the use of a lossy compressor, is a significant advance in the art.
  • One preferred lossy compression circuitry employs adaptive predictive pulse coded modulation (ADPCM) .
  • ADPCM adaptive predictive pulse coded modulation
  • the lossy compressor circuitry 28, in the illustrated embodiment of the invention provides for a linear predictor 28a of the motion vector field and includes additional circuitry for reducing the number of bits (the bandwidth) required to describe the predictor and prediction error of the motion vector field.
  • the linear predictor 28a predicts the current motion vector (marked X) using the four previously coded neighbors (marked 0) (previously available over lines 30).
  • additional circuitry (embodied in switch circuitry 28b) makes a decision regarding three possible choices: a) Reset the estimated motion vector to zero (the signals over line 188) and send it to the receiver; b) Reset the estimated motion vector (over line 28c) to the predicted value and send it to the receiver; or c) Quantize the prediction error (the difference between the original estimated motion vector and the predicted motion vector) in quantizer 28d and send the quantization index to the receiver.
  • the additional circuitry 28b makes this decision using an entropy constrained error measure, that is, a weighted combination of the used bits and error (for example, a measure equal to (aN + bE) where "a" and "b" are constants experimentally determined for the apparatus).
  • aN + bE a measure equal to (aN + bE) where "a" and "b" are constants experimentally determined for the apparatus).
  • the result that produces the smallest entropy constrained error is chosen and transmitted.
  • the encoding process starts by sending one bit of information about whether or not to reset the estimated motion vector to zero. If it is reset to zero, nothing more is sent for this motion vector.
  • the lossy compressor employs arithmetic coding for the first two steps and Huffman coding for the last step.
  • the output of the lossy compressor circuitry over lines 30, as noted above, is passed to the encoder 18.
  • those signals are employed by the error circuitry 14 for determining what the receiver would have seen, absent any errors in the channel, and thereby providing the mechanism for determining the prediction error signal, that is, the signal representing the difference between what the receiver would have predicted based upon the coded motion signal representation over lines 30, and the true image input.
  • the output of the lossy compressor over lines 30 is used by a reconstructor circuitry 34 for producing, at its output, a signal representative of the measure of motion displacement, the motion vectors, on lines 32.
  • the output of the reconstruction apparatus 34, over lines 36, is directed to a motion field interpolation circuitry 38 which operates in the spatial domain to associate with each picture element a motion displacement vector.
  • the motion field interpolator resolves that data so that there is associated with each picture element, a motion displacement vector.
  • the resulting output of the motion field interpolator, over lines 40, is designated the motion reconstruction signal.
  • the motion reconstruction signal is applied to a motion compensation apparatus 42 which forms part of an error reconstruction loop 43.
  • the error reconstruction loop includes a frame buffer 44, a lossy compression circuitry 46, and a reconstruction circuitry 48.
  • the inputs to the lossy compression circuitry 46, over lines 22 and 51 respectively, are the original input image for the current frame and the estimated receiver signal, that is, the signal which, absent any further data, the receiver will reconstruct and display.
  • the lossy compressor 46 provides the receiver with further encoded data, the error reconstruction signal, for reducing, and in principle eliminating, the difference between the original input image and the estimated receiver signal. That difference is coded to reduce its bandwidth and the resulting signal, the error reconstruction signal over line 52, is delivered to the channel encoder 18.
  • the lossy compressor 46 in the above referenced Ericsson patent is a two-dimensional block encoder which employs a gain/shape vector quantization; and the output of the block transform can be advantageously further reduced in bandwidth and encoded according to the processes described above in connection with the lossy compressor 28.
  • a hierarchical entropy encoded, lattice threshold quantization encoding method and apparatus are advantageously employed in implementing the lossy compressor 46.
  • the error reconstruction signal is also sent to the reconstruction apparatus 48 which provides an operation which is the inverse to that imposed by the lossy compressor 46. There results, therefore, at the output of the reconstruction apparatus 48, an error reconstruction image over lines 54.
  • the error reconstruction image is added to the expected output of the motion compensator, (which is the estimated receiver image over lines 51) and the resulting signal, an estimated previous receiver image (the predicted receiver image for the previous frame), is stored in the frame buffer 44.
  • the input to the frame buffer 44 is the estimated previous receiver image.
  • This receiver image which takes into account all data received by the receiver, corresponds to the reconstructed receiver image for a frame.
  • the image output from the frame buffer over lines 64 is the image which the motion compensation circuitry 42 modifies in accordance with the output of the motion field interpolator 38 over lines 40.
  • the output of motion compensator 42 thus represents the predicted receiver image as a result of reconstructing the output data from lossy compressor 28.
  • the data from the channel is decoded by a channel decoder circuitry 70 and the resulting receiver error reconstruction signal over lines 72 and receiver coded motion signal representation over lines 74 are delivered to reconstruction circuitry 76, motion compensator 99, and reconstruction circuitry 78 respectively.
  • the reconstruction circuitries 76 and 78 each provide for decoding the codes employed by the transmitter to effect the operations performed by reconstruction circuitries 48 and 34, respectively, of the transmitter, as described in more detail hereinafter.
  • the output of the error reconstruction circuitry 76 is delivered to a recovery loop 80 in which motion compensating signals over lines 82 are added to the error image representation over lines 84 to produce a reconstructed receiver signal over lines 86. That signal is delivered to a digital-to-analog circuitry 90 and from there to a monitor 92 for viewing.
  • Motion reconstruction signals are generated by a motion field interpolator 96 corresponding to the motion field interpolator 38 of the Figure 2.
  • the motion field interpolator provides a motion vector for each picture element of the image and hence allows the frame interpolator to accurately predict what the image would have been at any selected time between received frames.
  • the reconstructed receiver images over lines 86 are successively stored in a frame buffer 98 and are delivered to a motion compensator 99 which also receives signals from the motion field interpolator 96.
  • the output of the motion compensator representing the expected receiver image in the absence of an error correction, corresponds to the signal over lines 51 in the transmitter, and is delivered to the adder 100 for combination with the output of the error reconstruction circuitry over lines 84.
  • the transmitter and receiver circuitries of Figures 2 and 3 can be modified in a number of ways as described, for example, in United States Patents Nos. 4,727,422 and 4,816,914 referred to above. While these alternate embodiments of transmitter and receiver structure are applicable in different communications configurations, the invention described and claimed herein relating to the entropy coded, lattice threshold quantization encoding system is not dependent upon which of those particular transmitter configurations is employed and will therefore be described solely in connection with the typical transmitter and receiver configuration set forth hereinabove.
  • the motion field interpolator (38, 96) of the transmitter and receiver circuitries and the lossy compressor 28 of the transmitter circuitry are described in detail in Ericsson, U.S. Patent 4,849,810, issued July 18, 1989.
  • the lossy compressor 46 is modified by its use of the entropy encoded, lattice threshold quantization encoding method. Accordingly, the lossy compressor is described in detail herein. Further, the motion estimator and an adaptive filter which can be advantageously used in accordance with the present invention, are also described in detail in the above identified Ericsson, U.S. Patent 4,849,810.
  • the lossy compressor 46 receives as inputs the original uncoded signal over line 22 and signals representing the estimated receiver image over lines 51.
  • the lossy compressor 46 uses those signals for encoding the difference between them, and outputs the encoded error reconstruction signal over lines 52. This signal corrects for most errors not properly compensated for by the motion compensation system.
  • the estimated receiver image over line 51 (often referred to as the "warped" image) and the original uncoded image over lines 22 are decimated (that is, filtered and subsampled as described below) by decimation circuitries 502, 504, respectively, four times.
  • the image is subsampled by a factor of two both horizontally and vertically.
  • five levels of images for the luminance image are available at resolutions of, in the illustrated embodiment, 256x240, 128x120, 64x60, 32x30, and 16x15 picture elements for the luminance.
  • the set of images, at the different image resolutions, is commonly referred to as a "resolution pyramid."
  • the base of the pyramid is the full resolution image while the top of the pyramid is, in the illustrated embodiment, the 16x15 pixel image.
  • encoding of the image difference between the warped image and the original uncoded image is performed by an encoding circuitry 506 on a level by level basis, from the top level to the bottom level of the resolution pyramids.
  • the process terminates at that resolution when no additional bits are available for video transmission.
  • the system will typically reach the bottom or base level of 256x240 pixels while during a heavy motion the encoding may stop at the 128x120 level.
  • the apparatus will run out of transmission bits earlier in the pyramid or divide the available lists among the several pyramid levels.
  • encoding begins at the top level, that is, .the 16x15 image.
  • the 16x15 version of the warped image or a background image is used as the prediction. Recall that this corresponds to the image (decimated) that is created at the receiver absent any additional information.
  • this top level prediction is subtracted from the 16x15 decimated top level image of the original image.
  • the difference image, representing the error at that top level is quantized and the quantized information is directed to the encoder 18 for transmission to the receiver. Thereafter, the quantized difference image is added to the prediction image, at the 16x15 level, to form a 16x15 reconstructed image which the receiver will also create.
  • the prediction version of the image is formed in a different fashion.
  • the prediction is derived from the higher level reconstructed image and from the current level warped image as follows.
  • an interpolation error image is derived by interpolating the higher level warped image and subtracting it from the current level warped image.
  • the resulting warped interpolation error image thus essentially extracts the spatially higher frequencies of the warped image, that is, information not present in the higher level image.
  • the higher level reconstructed image is then interpolated to form an interpolated, reconstruction image at the current level.
  • the warped interpolation error image or the background image is selectively added to the interpolated reconstruction image to generate the prediction image.
  • the warped interpolation error image or the background image is used where it improves the prediction but not otherwise. This is decided on a block-by-block basis. and the decisions are transmitted to the receiver as "side" information.
  • the steps for generating the difference signal at this lower level are the same as those at the top level, that is, the current level prediction image is subtracted from the current level original image and that difference is quantized and transmitted to the receiver. Thereafter the quantized difference is added to the prediction image at that level to form a new reconstruction image. This procedure is repeated through the resolution pyramid until the bottom level is reached.
  • the reconstructed image at the bottom level is the output image at the level, and it is that image that is displayed by the decoder. That image is also used as described above to form a warped image for the next frame.
  • the warped image reconstruction at the transmitter is, as noted above, performed by the reconstruction circuity 48.
  • the predictions at the lower levels are still generated in the same manner; however, no coding, that is, no quantized difference information is sent to the receiver. Instead, the prediction at the lowest levels will be used directly as the output or reconstruction image at that level and as the error reconstruction image over lines 54 from reconstructor circuitry 48.
  • the resolution pyramid is formed, as noted above, by decimating four times, in this illustrated embodiment, the highest resolution level of the image.
  • each pair of pixels at a lower level are averaged to form a single pixel at an upper level.
  • the situation is the same both horizontally and vertically so that each higher level picture element is located at the center of a 2x2 pixel group of the lower level.
  • the coding method also provides for generating, using an interpolation procedure, the pixels at a lower level from a higher level.
  • the interpolation process is applied, for example, to the warped and reconstructed images to obtain images for processing at the next lower level and is effected by a bilinear interpolation.
  • the interpolation factors are 0.75 and 0.25.
  • arithmetic coding is employed for both coding of information for transmission from the lossy compressor 28 as well as, and as will be discussed in more detail below, the coding of scalar data from lossy compressor 46.
  • Arithmetic coding is well known to those skilled in the art. In particular, it can be applied advantageously to describing the locations of non-zero transform or other array variables. The symbol probabilities are changed depending upon previously transmitted values and the sequence position of the coefficient. Prestored probabilities are employed since on-line adaptation does not, in the experience of the inventor, provide significant improvement in this application.
  • the original and warped images have, at the top level, a resolution of 15x16 pixels for the luminance and 8x8 pixels for the chrominance, respectively.
  • Figure 6 describes the processing of the luminance component; and the processing of the chrominance component (not shown) can be similarly illustrated.
  • the prediction image consists of the top level warped image that was obtained originally by four decimations of the warped luminance and chrominance images, respectively.
  • the prediction error is generated by subtracting the prediction image 510 from the original, uncoded, top level decimated image 512.
  • the image differences over line 514 are quantized by a scalar quantizer 516 having a fixed step size.
  • the quantized information over line 518 is encoded separately for each component, the Y, the I, and the Q components, using the same arithmetic encoder 520 which is also employed for the motion vector transform coefficients.
  • Encoder 520 uses a Markov Model for encoding the non-zero data locations.
  • the encoder has sixteen states depending upon whether the already encoded four nearest neighbors corresponding to the four nearest neighbors illustrated in Figure 8 are zero or non-zero.
  • the non-zero values are encoded by a memory-less coder that encodes the eight bit quantization indices into the bit stream.
  • the quantized difference image is added to the prediction as noted above, and the result is the output or reconstruction image (over lines 522) at the top level.
  • the scalar quantizer 516 used in connection with the top level prediction error is a uniform quantizer having a dead-zone around zero. It codes the sign with a 1 bit and the magnitude with 7 bits in the illustrated embodiment.
  • the thresholds (T(i)) for the magnitude are located at:
  • a value of X where X is greater than T(k) but less than T(k+1) is assigned a quantizer index value of k and is reconstructed at the receiver as having a value R(k).
  • the quantizer is also symmetric around zero and sets all values with a magnitude less than T(l) equal to zero.
  • the prediction image is generated by combining the output image from the next higher level with the warped image of the same level. Then, the prediction error is formed by taking the difference of the original image at the current level. The difference image is coded using a lattice threshold quantizer and the quantized difference is added to the prediction to obtain a new output image at the current level.
  • the Y, I, and Q components are treated as three separate images.
  • the prediction image is generated by combining the warped image at the current level with the output and warped images from the next higher level.
  • the interpolation error of the warped image is generated using the warped image 524 at the current level and an interpolated version of the warped image from the next higher level (interpolated by circuitry 526). That interpolation error is thus the difference between the current level warped image and the same image that has been decimated and interpolated. As noted above, it contains the details o.f the warped image that were lost in the decimation to form the next higher level image.
  • the output image from the next higher level is then interpolated at interpolation circuitry 527 to obtain an image at the current level.
  • the warped interpolation error over line 528 or the background image is conditionally added by adder 530 to the interpolated output image to form the prediction. That is, for each block of 8x8 pixels, the squared error is determined between the original image stored at 532 and three possible predictions, that is, between the interpolated output image from the next higher level with and without the inclusion of the warped interpolation error, and also that with the inclusion of the background image.
  • the elimination of the warped interpolation error is equivalent to low pass filtering the warped image for the prediction.
  • This effective filtering process is performed in all blocks where it provides a significant decrease in the prediction error, that is, in those blocks wherein motion compensation was not successful.
  • the result of the filtering process termed “blurring,” is effected if the "blurred error,” multiplied by a weighting factor, such as 1.5 in the illustrated embodiment, is less than the error using the warped interpolation error.
  • the use of the background image is equivalent to a long term memory such that a representative of the non-moving object can be retained by the decoder even when these objects are briefly occluded. This background frame should only be used when a significant gain is achieved over other choices. Therefore, its weighting factor is 20% - 25% greater than the "blurring" weight.
  • the blur and background information generate a one or two bit (3 states) word for each 8x8 block. This is similar to the method used in motion vector field coding, and these two bits "answer" the following two questions: a) Does the system blur the current (8x8) block? b) If blurring is not preformed, should warped prediction or the background image be used?
  • a one indicates blurring and a zero indicates no blurring.
  • the information is encoded using an arithmetic coder 534 such as that noted earlier, and since each word contains only one bit, there is no need to encode the non-zero values once the "blur location map" has been encoded.
  • the particular arithmetic encoder 534 for the blur information uses six binary variables to select one of thirty-two states with corresponding probabilities.
  • the binary variables are the four previously encoded blur words for neighboring blocks at the same level and the blur word for the higher level neighbor, that is, the block at the next higher level that corresponds to the current block.
  • the encoder does not make explicit use of the fact that blurring at one level propagates to lower levels and instead this relationship is reflected in the probabilities for the various states having a non-zero higher level neighbor.
  • the prediction errors themselves are coded by the lattice quantizer 536.
  • the ⁇ , I, and Q components are treated as three separate images.
  • Each different image, generated for each level, is thus divided into blocks of 4x2 pixels.
  • Each block then becomes the "vector" and is coded by the lattice quantization by first determining a closest point and thereafter indexing the point as described in greater detail below.
  • the remaining levels can be encoded using the procedures applied at the (30x32) level, and using equivalents of elements 524, 526, 527, an adder 538, and elements 530, 534, and 536, but with the exception that the image data will be encoded, in the preferred embodiment, as described below.
  • the entropy coded, lattice threshold quantization described herein replaces the vector quantization described in Ericsson, U.S. Patent 4,849,810.
  • VQ vector quantization
  • lattice threshold quantization encoding we know, from Shannon Theory, that as the size and dimension of an optimally designed VQ codebook increases, its performance approaches the operational distortion-rate bound, without the need for entropy coding. Complexity in the VQ process, however, also grows without bound, and clever methods must be used to obtain good performance with only moderate complexity.
  • Lattice quantizers that are optimum in the Shannon sense were, prior to the invention herein, known for only a handful of sources and dimensions.
  • the hexagonal lattice is optimum for an independent, identically distributed uniform source in two dimensions. It has also been shown that a rectangular lattice works well for independent, identically distributed Laplacian sources. In general, one can show that lattices having high sphere packing densities work well for independent, identically distributed sources, provided the coding rate is high and an efficient entropy code is used for transmitting the indicies. Achieving good performance at low rates and for sources with memory, such as that described in this application, has been an elusive goal. Entropy Coded Lattice Threshold Quantization
  • an optimum performance over a wide range of data rates for a variety of data sources can be obtained for the first time using multiple dimension lattices.
  • the preferred encoding structure is termed the Entropy Coded Lattice Threshold Quantizer (EC-LTQ) illustrated in Figure 10.
  • EC-LTQ Entropy Coded Lattice Threshold Quantizer
  • the preferred embodiment of the invention uses an eight-dimensional lattice having the highest known sphere packing density, the E-8 lattice.
  • the vector quantizer referred to in Ericsson, U.S. Patent 4,849,810 is designed for a given (though possibly unmeasured) source probability density function (pdf) by successively optimizing during a training period, its encoder for its decoder, and its decoder for its encoder, etc., resulting in a locally (though not globally) optimal design that typically does not require further coding, such as entropy coding.
  • most lattice quantizers work best only for memoryless probability density functions.
  • the simplest one-dimensional lattice (the Z-l, or uniform quantizer) can not only approach, but can even attain the scalar quantizer's Gish-Pierce performance bound.
  • the excellent performance is obtained by using the centroid of each Voronoi region for the quantizer's reconstruction value, ( Figure 9) and by coupling the uniform threshold quantizer with an efficient entropy code (Figure 10).
  • the particular distortion-rate operating point is fixed by the quantizer's step size.
  • the Entropy Coded Lattice Threshold Quantizer of the present invention generalizes the uniform threshold quantizer to the vector case.
  • the E-8 lattice is used to encode the source, and a separate, optimal decoder is designed for it.
  • the encoder output is coded using a near-optimum entropy code.
  • the decoder consists of a countable set of centroids, one centroid for each lattice point's Voronoi region, given the source probability density function and distortion measure.
  • the lattice threshold quantizer's performance lies between the scalar quantizer's Gish-Pierce bound and Shannon's Operational Distortion-Rate Bound. While an entropy coded, lattice threshold quantizer can be built for any lattice, entropy coded lattice threshold quantization performance is relatively independent of the particular lattice employed, so even the simplest n-dimensional lattice, the n- dimensional cubic (or Z-n) lattice, can work about as well as the more complex, but denser, E-8 or 24- dimensional Leech lattices.
  • Lattice quantizers have no overload region, since even a source vector having arbitrarily large energy will still be located near a point in the lattice. A lattice quantizer can thus improve perceived performance of source coding systems, especially those needing occasional instantaneous high rates.
  • fast encoding methods for lattices exist and others are continually being sought, especially for lattices having high sphere packing densities (e.g., the E-8 and the 24-dimensional Leech lattice). These methods perform optimum encoding by locating the unique point in the lattice that is closest to a given source vector.
  • most fast encoding methods for vector quantization either perform a sub- optimum search of the vector quantization's finite codebook or, to attain optimum encoding, typically require substantially more memory than is needed to store the codebook itself.
  • entropy coded lattice threshold quantization is applied to a closed-loop pyramid image compression method
  • the entropy coded lattice threshold quantization techniques herein described are equally applicable to open-loop pyramid image compression, conventional subband and transform image and speech coding, and virtually any other compression technique using vector quantization.
  • the preferred embodiment of entropy coded lattice threshold quantization for hierarchical compression in accordance with the invention uses the E-8 lattice and the apparatus and method of the invention provide methods to efficiently implement entropy coded lattice threshold quantization for this lattice.
  • the E-8 lattice is composed of the union of one D-8 lattice with a second D-8 lattice translated by the vector 0.5 * (1 1 1 1 1 1 1 1).
  • the D-8 lattice is composed of all 8-dimensional vectors having integer components which sum to an even number. For example, (1 -3 0 0 0 0 0 0) is a point in D-8 lattice, while (1 1 1 0 0 0 0 0 0) is not. Since the translating vector also sums to an even number, the components of all points in E-8 lattice sum to an even number. Note also that the E-8 lattice includes the all zero vector.
  • All points in the E-8 lattice can be grouped by the value of their L-2 norms (sum of squares) into shells, where the shell number increases with the L-2 norm.
  • the all zero vector is the only lattice point having zero norm and it comprises Shell 0.
  • the probability density function is very "peaky" so that over 90% of the source vectors fall onto Shell 0. Of the remaining 10%, roughly 80% to 90% or more fall into Shells 1 and 2, making it worthwhile to code Shells 1 and 2 efficiently, but permitting simpler methods to be used for Shells 3 and up.
  • the entropy coded lattice threshold quantization process is performed in five steps: quantizing the source vector to the nearest lattice point, encoding the lattice point to a unique lattice index, entropy coding the index, decoding the index, and reconstructing the vector using a scaled version of the centroid associated with the lattice point.
  • the method implementation performs each step for the E-8 lattice.
  • the quantizer can identify the closest E-8 lattice point using one of several algorithms.
  • the quantizer first rounds the source vector to its nearest integer coordinates and then sums the coordinates to check whether the point is on the D-8 lattice. If the sum is even, this point is on the integer (not translated) D-8 lattice. If the sum is odd, then the closest lattice point (in the L-2 sense) is found by rounding the component having the largest roundoff error to the next nearest integer. Next, the quantizer finds a closest lattice point on the translated D-8 lattice. Each component of the source vector is truncated and increased by 0.5 and the resulting vector's components are summed.
  • the subtraction can also be obviated by comparing the sum to two.
  • Rn(new) 2 * [Rn(old)] - 1, while if a truncate-and-add-one-half coordinate changed, its corresponding influence on Rn is
  • each pair forms a 4x4 block of pixels.
  • Isolated ones at highest resolution level, that is, the 256x240 level, in this sequence are set to zero to limit "popping" effects. Arithmetic coding is used for coding this information.
  • the preferred implementation actually uses two sets of Huffman code tables, one table containing 120 codes for Shell 1, the other 800 codes for Shell 2. (One table from each set is selected by computing the ratio of the source vector's energy to the step size being used. This enables us to adaptively code the lattice points.) Finally, if the point is on Shell 3 or higher, we explicitly transmit the point's coordinates, usually by jointly Huffman encoding pairs of data.
  • the lattice point's energy can be computed and its shell identified.
  • Each eight dimensional lattice point corresponds to a two column by four row block of pixels.
  • Denoting the spatial relationships of the lattice points as: xO x4 xl x5 x2 x6 x3 x7 point pairs (xO, x4), (xl, x5), (x2, x6) and (x3, x7) are jointly coded. (If the pairs exceed some pre ⁇ determined dynamic range, they are individually coded.
  • the lattice point is reconstructed by taking the product of the lattice coordinate, step size and an attenuation factor, discussed later.
  • Shells 1 and 2 are coded using Huffman codes based upon computed probabilities.
  • a straightforward implementation would store 2400 8-dimensional centroids and 2400 Huffman codes for the 2400 points in the two shells.
  • considerable storage can be saved by noting symmetries between the points.
  • half the points are the negatives of the other half.
  • GN lattice point's 8-bit group number
  • one group number (having value zero) corresponds to the all zero Shell 0 lattice point
  • 120 each correspond to two of Shell l's 240 lattice points
  • the remaining 135 correspond to Shell 2's 2160 points. It is clear that a lattice point and its negative yield the same group number.
  • each of the 120 group numbers for Shell 1 corresponds to one lattice point and its negative.
  • each of the 135 group numbers for Shell 2 corresponds to exactly 16 lattice points, eight of which are negatives of the other eight.
  • Each group of eight are also mutually orthogonal, and therefore, the following method can be used for indexing them.
  • the vector is reconstructed by taking the product of the lattice coordinate or associated centroid, the step size and an attenuation factor.
  • the entropy coded lattice threshold quantization is required to operate over a variety of source distributions and bit rates.
  • entropy coded lattice threshold quantization is used for hierarchical pyramid video coding and the centroids in Shells 1 or 2 are used only at the three highest resolution levels of the pyramid. Yet the statistics of the source at each level differ. Also, at any one level, source probability density functions tend to change in variance as the step size changes. For example, if the input energy is large relative to the step size, the probability distribution over the lattice points is much broader than when the ratio is small. To combat these variations, several classes are defined and are determined by thresholding the ratio of the level's input energy to the square of the level's step size.
  • the apparatus stores a set of entropy code tables for each class. Also, the attenuation applied to the reconstruction value of a vector varies with the class, the level of the pyramid at which encoding is taking place, and the quantizer's shell. The reconstruction is thus given by:
  • Attenuation_table ( shell, level, class ) , where the attenuation factor varies with step size, as shown in Fig. 11.
  • These attenuation values are adjusted to limit "popping" artifacts in the reconstructed signal. In general, the cumulative attenuation is not great enough to move the reconstruction value outside the lattice point's Voronoi region.
  • the particular functional values established by the relationships defined here must be determined by teaming regression analyses with some heuristic constraints and perceptual experiments. They are implemented here as tables and piecewise linear functions, but could be generalized to continuous functional relationships.
  • the source data must be normalized by the step size to collapse its probability density function to a more stationary set of values.
  • the method employed to estimate proper step size is a bit allocation method. This method can also be used for the closed loop 5-level Hierarchical Pyramid vector quantization described in earlier patent Ericsson, U.S. 4,849,810. The method determines the number of bits to expend at each of the three higher resolution levels of the pyramid by specifying the step size used for that level. One level (3) is not coded and the last level (0) needs its own estimate, as described in the next subsection.
  • step sizes are selected such that the current level is coded with enough "bits" to match the prediction error energy of the next level.
  • the hierarchical coding structure is closed loop, so that all error energy remaining at the current level carries forward to the next level. The next level then attempts to code what is left to the level of prediction error in the succeeding level, etc.
  • the encoder builds a resolution pyramid of the current frame and creates predictions within the pyramid using motion compensation and background references.
  • Level 3 is never coded, since we can save some bits without any perceived loss in quality.
  • a method for computing the step size for level 0 can be estimated, in accordance with the invention, at level 0 in two steps. First an open loop estimate is made, followed by a closed loop estimate.
  • diff_bits(n-l) is the number of bits available less the number of bits used in frame (n-l), and h is a constant.
  • the diff_bits process is low pass filtered.
  • the system uses scalar quantization.
  • the warped interpolation error is selectively added to the interpolated output images from the top level and therefore, at this level, blur information is transmitted.
  • the coding will stop at the next level up (60x64 for luminance and 30x32 for chrominance). If the coding at this level generates more bits than the desired number of bits per frame, the coding is still performed and the frame rate will temporarily slow due to the large number of bits being generated. For the two bottom levels, the prediction is generated as usual and the blur information is transmitted; however, lattice threshold quantization is not performed.
  • the transmitted and coded data is decoded and the new frame generated.
  • the data representing the resolution pyramids is decoded by reconstruction circuitry 76 level by level from the top of. the pyramid down to the bottom.
  • the quantized difference image is decoded and added to the warped image at that level (in selective adder 100, which includes decimation circuitry).
  • the output image at the top level is reconstructed.
  • the lower levels are then reconstructed (by adder 100) by first forming the prediction using the transmitted blur information available over lines 84 from reconstruction circuitry 76 and then decoding the difference image and selectively adding it to the prediction image to form a new reconstructed image for that level. The process continues until the bottom level is reached and the bottom level image is transferred to the display frame buffer 98.
  • the arithmetic decoder which is used in a number of processes, that is, decoding the motion transform coefficients, the blur information, and the image information from lossy compressor 46, operates as is well known in the art. Since the non-zero locations were encoded using different probabilities depending upon which state the coder was in, the arithmetic decoder regenerates the state for each position in the location map as the decoding proceeds. The state, in combination with the encoded data, then determines whether or not a zero is used for each map location. Once the map indicating the location of non-zero values has been decoded, the B-bit values are decoded and incremented by one and placed in the appropriate positions in the map.
  • the process follows the inverse of the method used in lossy compressor 46.
  • the prediction at the top level is generated from the warped top level image.
  • the quantizer indices are decoded using the arithmetic decoder and the quantized difference image is reconstructed from the quantizer indices and is then added to the prediction image to give the top level output image (corresponding to adder 522a and the output over lines 522 of the transmitter).
  • the prediction is formed by selectively adding the warped interpolation error to the interpolated output image from the next higher level (corresponding to adder 530 in the transmitter).
  • That output image, and the warped image at the next output level, are interpolated to give images at the current level.
  • the blur information is decoded using the arithmetic decoder and then, in each 8x8 block of the interpolated higher level output for which the blur code is zero, the difference between the current level warped image and the interpolated higher level warped image is added (corresponding to adder 538 of the transmitter) .
  • the lattice threshold quantized information is decoded.
  • Each 4x4 block having a non-zero map bit is decoded by first determining the shell numbers of the two 4x2 lattice coded vectors inside. The centroid is then found for each lattice not lying in Shell 0 and scaled by the attenuation value. The output image is thereafter formed by adding this reconstructed difference to the prediction. (This corresponds to the operation of the adder 540 of the transmitter.) The output image from the bottom level in the final reconstructed image is then transferred to the display frame buffer as the final output image.
  • the illustrated lossy compressor 46 can be implemented in hardware, in software, or in a combination thereof.
  • the hierarchical entropy coded lattice threshold quantization encoding method and apparatus described hereinabove can also be advantageously employed in connection with the transmission of a sequence of images in which motion compensation is not provided.
  • the estimated receiver image over, for example, lines 51 in Figure 2 will be the reconstructed receiver image designated as the frame output of the frame buffer 44 over lines 64.
  • the input image over lines 22 and the estimated receiver image over lines 51 need not be individually decimated by lossy compressor 46 but can be, referring to Figure 7, input to a difference circuitry 720, the output of which, representing the error signal at the receiver, can then be input to a lossy compressor for decimation and hierarchical lattice threshold quantization encoding in accordance with the invention described hereinabove.
  • the motion compensator 99 and its related circuitry would similarly be eliminated if motion compensation were not employed.
  • the reconstruction circuitry 76 would be modified in accordance with the transmitter operation to reconstruct the error image representation over lines 84 when the circuitry of Figure 7 is employed.
  • the entropy coded, lattice threshold quantization can also be employed in coding other sequences of multidimensional data vectors having a uniform vector quantization.
  • the simplifications detailed above can be employed with such data, and would be modified to take account of the probability density functions for the vector data. Additions, subtractions, deletions, and other modifications of the preferred particular embodiments of the invention will be apparent to those skilled in the art and are within the scope of the following claims.

Landscapes

  • Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Compression Or Coding Systems Of Tv Signals (AREA)
  • Compression, Expansion, Code Conversion, And Decoders (AREA)

Abstract

Procédé et appareil de codage de données d'erreur entre trames dans un système de transmission d'images, et notamment dans un système de transmission d'images à compensation de mouvement, servant à transmettre une séquence de trames d'image d'un émetteur (8) à un récepteur (21). Le procédé et l'appareil utilisent une technique de quantification de seuil en treillis à codage par entropie hiérarchique (46) pour augmenter la compression des données relatives aux images transmises. Le procédé et l'appareil déciment (502) des données prévues entre trames et des données d'image non codées (504), et appliquent un codage par quantification de seuil en treillis à codage par entropie hiérarchique (506) aux structures pyramidales de données résultantes. On applique niveau par niveau un codage avec beaucoup de pertes, afin de produire la représentation sous forme de données codées des différences entre les données d'image prévues et l'image originale non codée. On peut utiliser ce procédé et cet appareil dans des systèmes transmettant une séquence de trames d'image (ou d'autres configurations de données, telles que celles relatives à un signal vocal) avec ou sans compensation de mouvement.
EP19910911119 1990-05-11 1991-05-10 A hierarchical entropy coded lattice threshold quantization encoding method and apparatus for image and video compression Withdrawn EP0527947A4 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US52260290A 1990-05-11 1990-05-11
US522602 1990-05-11

Publications (2)

Publication Number Publication Date
EP0527947A1 true EP0527947A1 (fr) 1993-02-24
EP0527947A4 EP0527947A4 (en) 1993-12-15

Family

ID=24081532

Family Applications (1)

Application Number Title Priority Date Filing Date
EP19910911119 Withdrawn EP0527947A4 (en) 1990-05-11 1991-05-10 A hierarchical entropy coded lattice threshold quantization encoding method and apparatus for image and video compression

Country Status (4)

Country Link
EP (1) EP0527947A4 (fr)
JP (1) JPH05509209A (fr)
CA (1) CA2082641A1 (fr)
WO (1) WO1991018477A1 (fr)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5134478A (en) * 1991-02-19 1992-07-28 Intel Corporation Method and apparatus for compressing and decompressing a digital video signal using predicted and error images
US8374446B2 (en) 2007-09-28 2013-02-12 Vsevolod Yurievich Mokrushin Encoding and decoding of digital signals based on compression of hierarchical pyramid
EP3459250A1 (fr) * 2016-05-16 2019-03-27 Numeri Ltd. Nouvel algorithme pyramidal destiné à la compression vidéo et à l'analyse vidéo

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US4853779A (en) * 1987-04-07 1989-08-01 Siemens Aktiengesellschaft Method for data reduction of digital image sequences
WO1991018478A1 (fr) * 1990-05-11 1991-11-28 Picturetel Corporation Procede et appareil de codage hierarchique utilisant des references de fond afin de permettre une communication efficiente de sequences d'images

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
GB8429879D0 (en) * 1984-11-27 1985-01-03 Rca Corp Signal processing apparatus
US4703350A (en) * 1985-06-03 1987-10-27 Picturetel Corporation Method and apparatus for efficiently communicating image sequences
US4727422A (en) * 1985-06-03 1988-02-23 Picturetel Corporation Method and apparatus for efficiently communicating image sequence having improved motion compensation
US4794455A (en) * 1987-01-07 1988-12-27 Pictel Corporation Method and apparatus employing adaptive filtering for efficiently communicating image sequences
US4816914A (en) * 1987-01-07 1989-03-28 Pictel Corporation Method and apparatus for efficiently encoding and decoding image sequences
NL8701261A (nl) * 1987-05-27 1988-12-16 Philips Nv Televisie-overdrachtsysteem met pyramidale kodeer/dekodeerschakeling.
US4849810A (en) * 1987-06-02 1989-07-18 Picturetel Corporation Hierarchial encoding method and apparatus for efficiently communicating image sequences

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US4853779A (en) * 1987-04-07 1989-08-01 Siemens Aktiengesellschaft Method for data reduction of digital image sequences
WO1991018478A1 (fr) * 1990-05-11 1991-11-28 Picturetel Corporation Procede et appareil de codage hierarchique utilisant des references de fond afin de permettre une communication efficiente de sequences d'images

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
1989 INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING vol. 3, 23 May 1989, GLASGOW (SCOTLAND) pages 1743 - 1746 D.G. JEONG ET AL. 'Lattice Vector Quantization for Image Coding' *
See also references of WO9118477A1 *

Also Published As

Publication number Publication date
CA2082641A1 (fr) 1991-11-12
EP0527947A4 (en) 1993-12-15
WO1991018477A1 (fr) 1991-11-28
JPH05509209A (ja) 1993-12-16

Similar Documents

Publication Publication Date Title
US5150209A (en) Hierarchical entropy coded lattice threshold quantization encoding method and apparatus for image and video compression
US5155594A (en) Hierarchical encoding method and apparatus employing background references for efficiently communicating image sequences
US4851906A (en) Data compression using orthogonal transform and vector quantization
EP0705035B1 (fr) Codage de données représentées par un réseau multi-dimensionnel
US5576767A (en) Interframe video encoding and decoding system
Fränti et al. Compression of digital images by block truncation coding: a survey
US4816914A (en) Method and apparatus for efficiently encoding and decoding image sequences
US5610657A (en) Video compression using an iterative error data coding method
US5909513A (en) Bit allocation for sequence image compression
JPH03136595A (ja) デイジタル動画像の記憶または伝送に用いるビデオ画像データの処理方法
Fowler et al. Image coding using differential vector quantization image coding using differential vector quantization
EP0527947A1 (fr) Procede et appareil de decodage par quantification de seuil en treillis a codage par entropie hierarchique pour compression image et video
EP0527952B1 (fr) Procede et appareil de codage hierarchique utilisant des references de fond afin de permettre une communication efficiente de sequences d'images
Yang et al. Motion-compensated wavelet transform coder for very low bit-rate visual telephony
KR970003105B1 (ko) 적응적인 패턴 양자화를 이용한 영상 부호화 장치 및 양자화 패턴 선정 방법
Vasconcelos Library-based image coding using vector quantization of the prediction space
Sun et al. Predictive vector quantization based upon first-order Markov model for image transmission
Fowler An algorithm for image compression using differential vector quantization
Fowler Jr et al. Design of a Real-Time Video-Compression System Using Differential Vector Quantization
Gharavi‐Alkhansari et al. Causal fractal compression of video sequences using matching pursuit
Shen et al. A Fixed Quality, Variable Rate Finite State Vector Quantizer For Encoding Video Over Packet Networks
Liang Adaptive vector quantization on image sequence coding
Refai Real-time transmission of digital video
Yang et al. Wavelet transform coding for MCP error images using quadtree decomposition and ECVQ
KR19980050957A (ko) 영상 신호 벡터 양자화기를 위한 다중 부호어 전송 방법

Legal Events

Date Code Title Description
PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

17P Request for examination filed

Effective date: 19921208

AK Designated contracting states

Kind code of ref document: A1

Designated state(s): BE CH DE ES FR GB IT LI SE

RIN1 Information on inventor provided before grant (corrected)

Inventor name: BERNSTEIN, JEFFREY

Inventor name: GIROD, BERND

Inventor name: YUAN, XIANCHENG

Inventor name: BAKER, RICHARD, L.

Inventor name: THOMPSON, EDMUND

A4 Supplementary search report drawn up and despatched

Effective date: 19931022

AK Designated contracting states

Kind code of ref document: A4

Designated state(s): BE CH DE ES FR GB IT LI SE

17Q First examination report despatched

Effective date: 19950804

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: THE APPLICATION IS DEEMED TO BE WITHDRAWN

18D Application deemed to be withdrawn

Effective date: 19960215