JP2007189704A - Recognition device for image - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To suppress generation of a retransmission mode and picture freezing as much as possible while transmitting a moving picture through super-low-rate transmissions. <P>SOLUTION: The super-low-rate image transmission can be performed by providing an encoding control section 21, a property memory 22, a property prediction section 23, and an area extraction/recognition processing section 25, and totally using an adaptive control over change of syntax, substitution of code words, prediction of encoding properties of a current frame based upon past encoding properties, and decision making of properties, object area extraction and quantization-by-area control by movements and models, control over the significant number of conversion coefficients corresponding to a use mode, a transmission rate, and a movement generation quantity, etc. Further, constitution conforming to the current image compression standards (H. 261) is embodied with simple constitution with an attached protocol converter. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to an image recognition apparatus suitable for image signal encoding processing in which an image signal is encoded and transmitted as an encoded bit stream compressed using a variable length code.

  In recent years, the use of multimedia in information communication terminals has been rapidly progressing, and it is essential for future business development to create various added values based on digital data transmission for telephones. In particular, in PHS (personal handy phone system), transmission of not only conventional voice and text information but also natural still images as well as moving images has been studied in earnest, and the use of multimedia for phones is very close. It is expected to be natural in the future.

  However, when transmitting such an image signal, in the uncompressed state, how to convert a TV image having a signal amount corresponding to about 166 Mbps to a PHS (32 kbps or 48 kbps among 384 kbps) or the transmission rate of an existing telephone line. There is a problem of whether to transmit at (about 10 kbps).

  Here, a moving image compression technique used for a conventional TV conference or video CD is required. However, even in this standard, for example, the current video conference standard ITU-T / H. 261 targets 64 kbps to 2 Mbps, and ISO / MPEG1 targets about 1.5 Mbps, and does not support ultra-low rate transmission of 64 kbps or less. It should be noted that H.264 recommended in November 1995. The H.263 standard targets 4.8 kbps to 64 kbps, but since the contents have not been disclosed, the system cannot be recognized. In the standard, H.264 The connectivity with H.261 is not compensated.

  Furthermore, such moving image compression techniques are mainly intended for wired systems. For example, the level of data transmission errors (−2 to −3 of 10) occurring in wireless communication in mobile communication is used. In the wireless system, a method called demand refresh using a retransmission request by ARQ (automatic repeat request) is usually used. However, when such ARQ is used, it is inevitable that the picture freezes until the retransmission of the image information is completed on the receiving side. Accordingly, there is a problem that the picture freezes in a place where the radio wave condition is bad, and the moving picture transmission is substantially impossible.

  On the other hand, for still image transmission, a method of dividing data into groups according to importance and rate control has been proposed, but it has not been put into practical use. In addition, an example of dealing with transmission errors by changing the data transmission order and error correction code at a transmission rate of about 384 kbps has been reported, but there has not yet been reported in ultra-low rate wireless transmission of 64 kbps or less.

  The present invention has been made in view of the above circumstances, and its purpose is to enable transmission of moving images at ultra-low rate transmissions such as PHS and digital cellular phones as a function suitable for image signal encoding processing. However, an object of the present invention is to provide an image recognition apparatus capable of suppressing the occurrence of a data retransmission mode and a picture freeze as much as possible.

  With respect to the image recognition apparatus, by setting the model base mode as the encoding control mode, the position information and the color information corresponding to the movement are encoded based on the three-dimensional model. By transmitting as data, it is possible to perform coding control with the amount of code generation suppressed as much as possible. In this case, in claim 2, distance information can be calculated by using a two-dimensional template, and in claim 3, the position and orientation of the three-dimensional model is estimated from the position data of the feature region based on the two-dimensional motion information. Will be able to.

  According to the present invention, when there is insufficient information for determining the mode by analyzing the image signal, the mode is automatically shifted to a mode having a higher probability based on a preset transition probability value. To be able to. In this case, in claim 5, the transition probability value is updated and set so as to suppress the error when the result of transition based on the transition probability value is evaluated and the error value becomes large. It becomes possible to achieve transition between modes suitable for the form. According to the sixth aspect of the present invention, the above-described evaluation can be surely performed by a two-dimensional template, a color analysis result, or a speed / acceleration evaluation.

  According to the seventh aspect of the present invention, it is possible to instruct and guide the setting information to the user through the human interface so as to adapt to the incidental situation corresponding to the mode setting state by the mode control means, It will be possible to contribute to the reduction.

  According to the present invention, the target area existing in the screen is extracted by the target area extracting means, and the target area information is added as attribute data to the image signal and transmitted by the encoding means. The amount can be reduced. In this case, according to the ninth aspect, the target area can be extracted by detecting the generation area of the motion vector as the attribute information specified in block units.

  Further, in claim 10, since the motion vector generation region is detected by performing analysis in the time direction over three frames or more, the target region can be reliably detected. Since the block including the maximum motion vector is detected and the region including the block is estimated as the main region, the target region can be extracted with high probability. Further, in the twelfth aspect, when the maximum motion vector is detected, the target region can be extracted after subtracting the panning vector calculated by the panning vector calculation means.

  Further, according to the thirteenth aspect, when extracting the target region, the feature amount is calculated from the block pattern composed of the motion blocks, and scaling is performed according to the presence / absence of the template information. To be able to do that. In this case, in claim 14, the template information can be model-based transmitted as attribute information in units of blocks. Further, according to the fifteenth aspect, the movement can be easily detected by analyzing the pattern in which the object in the screen moves two-dimensionally.

  According to the sixteenth aspect, in a person mode under a limited environment, a two-dimensional template can be generated by performing perspective transformation of a three-dimensional model. According to the seventeenth aspect, the category of the target area is determined based on the control state according to the mode information, and thereby, it is possible to shift to the determination condition for performing the state transition. Further, in claim 18, the category of the target area can be determined according to the condition of the state transition in the mode control. Furthermore, according to the nineteenth aspect, the category of the target area can be determined based on attribute information set in units of blocks.

  According to the present invention, when the forced intra frame is transmitted by the target region extraction unit, the region extraction process is performed, and the background region information estimated from the target region is specified and transmitted. Decoding processing can be performed accurately using this information on the decoding device side. In this case, in claim 21, by storing the background information based on the information designated on the decoding device side, the information on the background area is stored when the information amount of the target area portion increases. It becomes possible to reproduce information with reference to the information, and it is possible to maintain the reproduction of the moving image while preventing an increase in the amount of generated code. Further, according to the twenty-second aspect of the present invention, at the time of encoding processing, by adding data to be combined with the person information being transmitted to the information specified as the background area information, the decoding device side reproduces the information by the person information of the target area portion. By combining the person image to be processed and the background area information decoded in the past frame, it is possible to reproduce an image to which the background area image is added.

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment in which the present invention is applied to a configuration assuming a car phone will be described below with reference to the drawings.
FIG. 1 shows a functional block configuration of an encoder 1 as an encoding device, and FIG. 2 shows a functional block configuration of a decoder 2 as a decoding device. The overall configuration will be schematically described below with reference to FIG. 1 and FIG.

  In the encoder 1 shown in FIG. 1, a camera 3 as an imaging unit captures an object and outputs image data as a video source as a normal analog signal. In this case, the video source is not limited to the image source such as the camera 3, and a video signal stored in advance can be used. A video source based on image data photographed by the camera 3 is input to the A / D converter 5 via the changeover switch 4 and converted into a digital signal. The A / D converter 5 is connected to the orthogonal transform unit 7 via the changeover switch 6. In this case, the A / D converter 5 is also connected to the orthogonal transform unit 7 through the subtracter 8 and the changeover switch 6. Then, whether or not the subtracter 8 is interposed is controlled by switching setting of the changeover switch 6.

  The orthogonal transform unit 7 is an H.264 format. As in the H.261 standard, the original image or the prediction difference image is subjected to two-dimensional orthogonal transformation (in this embodiment, according to the H.261 standard, discrete cosine transformation (DCT)) along the CIF format. The converted output is output to the quantizing unit 9. The quantizing unit 9 quantizes and transforms the transform coefficient data given from the orthogonal transform unit 7 with 9-bit accuracy by linear quantization or quantization with a dead zone, and outputs the quantized transform data to the channel encoding unit 10. . The communication path encoding unit 10 converts the quantized image signal from the quantization unit 9 into an encoded bit stream in accordance with a predetermined conversion rule. The data is output to a communication path 12 such as a wireless path.

  The inverse quantization unit 13 is supplied with the quantized signal converted by the quantizing unit 9. The inverse quantizing unit 13 converts the image signal quantized by the quantizing unit 9 back to digital data before quantization and reverses it. The data is output to the conversion unit 14. The inverse conversion unit 14 converts the digital data into analog data, restores the image data so as to be substantially the same as the image data obtained from the video source, and outputs the restored data to the prediction memory 16 via the adder 15. The prediction memory 16 is for predicting the image data in the next frame based on the image data fetched from the video source, and outputs it as a subtraction signal to the subtracter 8 via the loop filter 17 and the changeover switch 18. It has become.

  The loop filter 17 functions as a spatial low-pass filter that performs smoothing to suppress a sudden change in color or brightness between adjacent blocks. The output of the loop filter 17 is changed from the changeover switch 18 to another changeover switch. 19 is provided as an addition signal to the adder 15 via 19. Further, the output of the prediction memory 16 is supplied to the motion detection unit 20, and the motion detection unit 20 generates a frame from both data of the video source image data supplied from the A / D converter 5. A signal is given to the quantizing unit 9 and a signal is given to the communication path coding unit 10 so as to compensate for the motion between them.

  The encoding control unit 21 executes various types of control of the encoding process. The conversion information is given from the orthogonal transformation unit 7 and the data such as the buffer usage rate is given from the FIFO buffer 11. Based on these data, the adaptive quantization control for setting the level of the quantization transformation of the quantization unit 9 is performed, and the adaptive rate control is performed. The attribute memory 22 that stores the attribute data of the image signal obtained as a result of encoding is configured to receive attribute data to be stored from the encoding control unit 21, and the attribute prediction unit 23 receives the attribute memory. The attribute data currently being encoded is predicted and given to the encoding control unit 21 based on the past attribute data stored in 22. As described above, H. An encoding processing unit 24 based on the H.261 standard is configured.

  Next, a characteristic configuration added in the present embodiment will be described. That is, the region extraction / recognition processing unit 25 is configured as follows. The target area extraction unit 26 executes target area extraction processing, which will be described later, based on the motion vector information given from the motion detection unit 20, and outputs information to the template database 27. The 2D (two-dimensional) template matching unit 28 performs matching processing based on the data from the template database 27 and outputs the processing result to the target region extraction unit 26 and the communication path encoding unit 10.

  The model base prediction unit 29 calls data from the three-dimensional shape database 30 and performs image reproduction processing in the model base mode based on the data stored in the person memory 31 and the background memory 32. A subtracting input is provided to the subtracter 8 via the conversion unit 10 and the changeover switch 18. The person memory 31 and the background memory 32 are configured to receive person data and background data to be stored from the target area extraction unit 26.

  Further, a mode control unit 33 for performing mode control of the encoder 1 is connected to the encoder 1, and detailed contents of various mode controls are implemented as described later. . The mode controller 33 is connected to a human interface 34 for exchanging information with the user. Specifically, the human interface 34 includes a display unit, a voice input / output unit, an operation input unit, and the like, and is controlled to perform mode control quickly and appropriately. The mode control unit 33 is also connected to the decoder 2 side, and as described in the section of operation, the mode control of each unit of the decoder 2 is similarly performed. The decoder 2 can also use the human interface 34.

  The person image information can be registered and stored in the person memory 31 as personal identification information through the human interface 34, and the personal identification information is input via the human interface 34 at an appropriate timing. The information is called from the person memory 31 based on the received information, and can be displayed as registered personal identification information even in cases other than communication.

  Next, in the decoder 2 shown in FIG. 1, the channel decoding unit 35 that receives the encoded bit stream of the image signal data transmitted via the transmission path is configured with a hierarchical encoding attribute (macroblock). Type) and the like. It is connected to a decryption processing unit 36 having a basic configuration conforming to the H.261 standard. The error correction function includes a pattern / signal level error correction unit 37, a recognition level processing unit 38, and a memory 39 as a storage unit for storing various data. Further, in the decoding process, the above-described mode control unit 33 performs control by exchanging signals with the above-described units, and the finally decoded image signal is changed to the changeover switch 40 and the D / A. The signal is output via the converter 41.

  In the above-described communication path decoding unit 35, the encoded bit stream data of the image signal transmitted via the external communication path or the transmission path 12 such as the air is input to the FIFO buffer 42, and data of at least one frame. Is accumulated. The encoded bit stream data stored in the FIFO buffer 42 is subjected to a grammatical interpretation of the code level in the parser 43. The fuzzy matching unit 44 globally searches the bit stream data stored in the FIFO buffer 42, specifies the PSC and GBSC described by the specific code by fuzzy matching processing, and also uses the ESC, EOB and the like are searched and specified, their positions are stored in the storage unit 45, and used when an error correction function is executed as will be described later.

  The parser 43 searches for other codes and data and interprets the grammar based on the position of the specific code searched by the fuzzy matching unit 44. In this case, the variable length code (VLC) ) Is referred to. The error determination unit 47 determines whether or not there is an error when performing a grammatical interpretation in the parser 43 based on the position of the specific code specified by the fuzzy matching unit 44. Are corrected at the code level and output. The parser 43 outputs mode information obtained by the decoding process of the bit stream data to the mode control unit 33.

  In the decoding processing unit 36, the inverse quantization unit 48 is supplied with a decoded output from the parser 43, and performs inverse quantization conversion on the quantized image signal data and outputs it to the inverse conversion unit 49. To do. The inverse transform unit 49 outputs orthogonal transform data corresponding to the frequency signal included in the image signal as decoded data by performing inverse transform to data corresponding to the position signal.

  The output of the inverse conversion unit 49 is input to the error determination / correction unit 50 via the path input to the error determination / correction unit 50 of the pattern / signal level error correction unit 37 and the changeover switch 51 and the adder 52. Routes are provided. The prediction memory 53 is supplied with data from the error determination / correction unit 50 and the motion compensation unit 54, and is connected to the addition input of the adder 52 via the loop filter 55 and the changeover switch 56.

  In the pattern / signal level error correction unit 37, the error determination / correction unit 50 provides data to the frame memory 57, and estimated image data is provided via the pixel value estimation unit 58 and the image estimation unit 59. Yes. The motion vector estimation unit 60 obtains data from the memory 39 and outputs the estimated motion vector data to the motion compensation unit 54. Also, the attribute estimation unit 61 reads past attribute data stored in the memory 39 and gives it to the decoding control unit 62 of the decoding processing unit 36 to provide information on the decoding operations of the inverse quantization unit 48 and the inverse transform unit 49. give.

The memory 39 includes a storage unit including a mode information unit 39a, a region information unit 39b, a pattern attribute unit 39c, a 2D (two-dimensional) motion vector unit 39d, and a personal identification information unit 39e.
In the recognition level processing unit 38, the model base prediction unit 63 is provided with 3D (three-dimensional) shape data 64a from the three-dimensional shape database 64, and with human image data from the person image database 65. Prediction is performed and data is output to the error determination / correction unit 50. The person image database 65 and the three-dimensional shape database 64 are given data from the memory 39.

  The background memory 66 as background information storage means and the person memory 67 as person information storage means are configured to be able to store background information and person information in the image signal output to the D / A converter 41. Data is output to the frame memory 57, and data is exchanged with the person image database 65. The region reproduction unit 68 is given a 2D (two-dimensional) template 69a selected in the template database 69 based on data from the memory 39, performs region reproduction, and outputs it to the frame memory 57.

  Next, the operation of this embodiment will be described with reference to FIGS. In the following description, [A] the operation content of the encoding process of the image signal in the encoder 1 will be described by the encoder, and [B] the encoded bit in the decoder 2 by the decoder. The operation content of the stream decoding process will be described. In the [A] encoder, various methods for reducing the code amount in the encoding process are listed by item (sections [A-1] to [A-4]). The explanation of the operation will be described mainly with reference to the flowcharts of the encoding processing program (FIGS. 5 to 7) (section [A-5]). [B] In the decoder, the decoding error in the decoding processing process Various methods of detection and correction are listed by item (sections [B-1] to [B-8]), and an explanation of the operation is described with a focus on flowcharts of decoding processing programs (FIGS. 8 to 11). (Section [B-9]).

[A] Encoder [A-1] Basic Operation Regarding the operation of the encoder 1, first, the basic operation will be described. As a basic operation of the encoder 1, image data captured by the camera 3 is captured, the image signal is encoded, and the encoded bit stream data is output to the communication path 12. In this case, the encoding process for the first frame and the encoding process for the second and subsequent frames are performed separately.

  First, in the first frame, a video source for supplying a scene including an object photographed by the camera 3 is converted into a digital signal by the A / D conversion unit 5, and then the H.264 is converted. Intraframe coding (INTRA) is executed in units of macroblocks according to the CIF format in accordance with the H.261 standard, and this is transmitted as an INTRA picture.

  In the second and subsequent frames, only the motion compensated prediction difference for the first frame is subjected to a two-dimensional orthogonal transform (here, DCT (discrete cosine transform) compliant with H.261) and then buffered by the encoding control unit 21. After being subjected to adaptive quantization in accordance with the quantity, coding rate, and other state values of the encoder 1, it is temporarily stored in the output FIFO buffer 11 via the channel encoder 10. The output FIFO buffer 11 transmits encoded bit streams having different amounts of information in units of frames to a communication path in accordance with a predetermined encoding rate.

  Note that the encoded bit stream is converted into data having a four-layer structure of PICTURE / GOB / MBK / BLK (see FIG. 30) in accordance with a CIF (common intermediate format) standard, as will be described later. It is generated with an attribute header attached to each layer for a set of variable length codes representing a sequence of DCT significant coefficients in the BLK (block) layer. On the other hand, the motion compensated prediction frame is generated as an output of the frame addition loop of FIG. 1, and is compared with the original image to send new difference information to the encoding mechanism.

  Now, in the above-described encoding process, the encoder 1 of the present embodiment performs an extremely low bit rate encoding process by performing the following unique processing.

  In other words, in this embodiment, H.264, which is an existing video compression international standard for video conferences and videophones. Based on the premise based on the H.261 standard, a mode control function and a region extraction function are added as unique processing steps. Under such a configuration, processing such as encoding control and syntax change is executed at each level of signal, pattern, code, and recognition, and finally an ultra-low rate of about 64 kbps to 4.8 kbps. The image encoding is realized. A conceptual correlation between these control contents is shown in FIG.

  In the following description, [A-2] mode control, [A-3] target area extraction, [A-4] encoding control operation, and [A-5] encoding operation program are described. The principle and method of the encoding method will be described by classifying into large items and further classifying each large item into smaller items.

[A-2] Mode Control FIG. 3 conceptually shows the contents of the ultra-low encoding rate in this embodiment. For the mode control performed in this embodiment, for example, (1) Usage environment, Mode control is performed in each of the three categories of (2) subject and (3) encoding control. In this case, the progress of the mode in each mode control is not always performed independently. For example, encoding may be performed using two or more modes simultaneously in each category.

  These transition relationships are governed by the states of the encoder 1 and the decoder 2, that is, the buffer amount of the FIFO buffer 11, the currently set mode state, the coding rate or the amount of motion generation, and the determination information. There are two cases: the case where it is done and the case where it depends on the stochastic state transition. In the following, the operation of mode control based on the difference between these categories (1) to (3) and state transition will be described.

(1) Use environment As for the mode control of the use environment, as shown below, considering the fixed and movable modes corresponding to each location of the automobile, indoors, and outdoors, (a) In-car fixed mode, ( It can be classified into six modes: b) in-car movable mode, (c) indoor fixed mode, (d) indoor movable mode, (e) outdoor fixed mode, and (f) outdoor movable mode.

(A) In-car fixed mode This is used to set the rear seat usage environment that is most conceivable for an automobile videophone, and this limits various conditions based on its positional relationship, for example, as shown in FIGS. There are features that can. This is summarized as follows.

1) The part of the background inside the car can be fixed substantially, and the scenery outside the car seen from the window becomes a moving area while driving.
2) The distance L from the camera 3 to the person P can be limited (see FIG. 12). As a practical value, the distance L can be set within a range of 1 m, for example.
3) As will be described in the next subject mode section, in this use environment, the subject is mostly a person mode centered on a person (see FIG. 13), and the area ratio is the largest in the result of moving region extraction. The probability that the region is a person is very high.
4) When using the rear seat as a target, the number of persons P to be photographed is most likely to be one.
5) In the case of a passenger car, the number of persons P that can enter the field of view of the camera 3 is considered to be about four at maximum.

(B) In-car movable mode In this mode, in addition to the person mode, a landscape mode as a subject mode is more likely to be used.
(C) Indoor fixed mode Although it is almost the same as the automobile fixed mode, the scenery outside the window is likely to be fixed, which is advantageous for reducing the amount of information.
(D) Indoor movable mode In addition to the portrait mode, the landscape mode is likely to be used.
(E) Outdoor fixed mode This mode can be applied to traffic monitoring, security, pet monitoring, and the like. In this case, since the camera 3 is fixedly installed, wired transmission is possible, and the transmission error rate is hardly lowered.
(F) Outdoor movable mode This is the harshest condition for use environment. In this case, since it is movable, it is a precondition to perform wireless ultra-low bit rate transmission. In addition, when using a portable device, there is a high possibility that the landscape mode is used in addition to the portrait mode. In addition, it is assumed that the vibration of the camera 3 increases due to camera shake or the like during a call.

(2) Subject In this case, the four encoding modes are switched and set as follows according to the encoding target and application.

(A) Person mode This person mode is set to a default mode used with the highest priority in normal communication. In this person mode, as shown in FIG. 14, mode switching as described below is further performed in accordance with the background and usage environment. In this case, the mode switching condition is normally set and switched automatically based on the extraction result of the target region and the distance determination.
A) Head mode (when there is only one person P)
B) Upper body mode (when there is only one person P)
C) Whole body mode (when there is only one person P)
D) Multi-person mode For example, in each use environment shown in FIG. 14, (a) the head mode of a) in the car, (b) the upper body mode in the indoors, and (c) outdoors. It is set to the upper body mode of a) or the whole body mode of c). In addition, in the multiple person mode of D), a particularly detailed template (see FIG. 25, see the description of the template to be described later) is not prepared, and it is assumed that the scene is a transitional short time. It is left to other control according to the amount of encoded information generated.

(B) Landscape mode For example, in a running car, the subject is often a person at the time of conversation, and the mode determination tree has a high probability of setting the person mode. On the other hand, if the user intentionally wants to transmit the scenery or things outside the vehicle as an image, encoding is performed based on the amount of information generated due to the presence or absence of movement and the fineness of texture and the presence or absence of people. Change mode to quasi-video or still image mode.

(C) Still Object Mode When a stationary object is a target, the amount of generated encoded information is relatively small, and therefore a normal moving image mode can be selected as the encoding mode.
(D) Drawing / character mode The high-resolution still image mode can be selected as the encoding mode.

(3) Coding control (a) Image center priority mode This mode is set when there is no foreseeing information for the object in the screen and there is no manual setting mode prior to this. The That is, as a situation, it is assumed that the camera 3 is pointed at an object that is simply desired to be photographed. In this case, normally, when the object is positioned at the center of the image by directing the camera 3, it can be assumed that the region to be watched is also a portion close to the center of the image (that is, the camera 3 with respect to the center of the object). (It can be assumed that the horizontal and vertical angles α and β of the posture are close to zero).

  Therefore, an encoding control operation is performed in which a bit amount is preferentially allocated near the center of the image and encoded, and the allocated amount is gradually reduced toward the periphery. Specifically, this is performed by [A-3] coding-by-region quantization control to be described later, use of background memory, control of the maximum number of significant transform coefficients, and the like. In this mode, conditions for shifting to another mode are set corresponding to the following three cases.

1) When the movement of the target area is detected If the target area is extracted, the process shifts to the target area priority mode. Here, since the target area cannot be tracked in the following case, the area centroid mode in the target area priority mode is selected.
A) When 2D (two-dimensional) template is not fixed
B) When 2D template adaptation is not complete
C) Landscape mode without a 2D template (Note that 2D templates will be described later.)

2) When the subject mode is the person mode In the target area priority mode and the person mode, the 2D template is suitable unless it deviates from the front image. In this case, identification of each part such as the head, eyes, and mouth is further started.

3) Object category identification based on target region extraction and motion analysis results When there is no motion information or region information, the screen center region is preferentially encoded in detail. In this case, giving priority to the screen center area is equivalent to considering a block within a range of a value calculated as an average distance Dav as a certain distance from the screen center as a target area, as shown in FIG.

  Here, in calculating the average distance Dav, the Eucrid distance D (n) shown in (a) or the absolute value distance D (n) shown in (b) is used as a distance scale. As shown in the following equation (1), the average value Dav of the distance D (n) [in this case, the absolute value distance is used as a reference] for the block (vertical 18 × 22 horizontal) set on the screen. As a result of the calculation, the average distance Dav is about 12, and the hatched area indicated by the oblique rectangle in FIG.

ただし、Ｎ＝２２×１８＝３９６、ｄｄ（ｉ，ｊ）は、絶対値距離では、
ｄｄ（ｉ，ｊ）＝｜ｉ−１１｜＋｜ｊ−９｜
となる。

However, N = 22 × 18 = 396, dd (i, j) is an absolute value distance,
dd (i, j) = | i-11 | + | j-9 |
It becomes.

(B) Target area priority mode 1) Background memory use mode [A-2] Based on area information obtained by area extraction of the target area, which will be described later, in the person mode, a background image prepared in advance or the first In this mode, the background information stored in the background memory 66 is used using the background image transmitted in the frame, and is controlled as follows.
A) Do not transmit any background information after that.
B) Although background information is transmitted, the background change information is periodically distributed and the background is gradually updated.
2) Region center-of-gravity mode In this region center-of-gravity mode, the image center in the above-described image center mode is considered to be the center of gravity of the target region, and the region-by-region quantization control is driven. Note that the region-by-region quantization control will be described later in [A-3] coding control, and will be omitted here.
3) Target area follow-up mode This mode allows the above-described 2) area center-of-gravity mode to be applied by detecting the position of the center of gravity of the target area that changes every moment when the target area moves on the image plane. Mode.

(C) Motion priority mode In this mode, the frame rate is not reduced even when the number of motion blocks or the absolute value of motion increases. Here, the intensity of movement is indicated by an index AM (see Expression (10)) indicating the intensity of movement in the description of adaptive control according to the information generation amount in the section [A-3] encoding control described later. By using the determination, the target area priority mode, the screen center priority mode, the model base mode described later, or the like is driven to control the transmission rate.

(D) Quasi-moving picture (INTRA picture) mode In this mode, in order to reduce the amount of information, the transmission rate is maintained while maintaining the image quality by lowering the frame rate. In this case, depending on the amount of motion, compression using the correlation between frames may be difficult. In this case, only the INTRA picture is transmitted.

(E) Still Image Transmission Mode This is assumed to be the following two cases.
1) Case where coding control fails In this case, first, a forced intra frame is newly transmitted as an image signal freezed by the encoder 1 side. Then, after this is completed, the mode is a mode for moving to the moving image transmission mode and restarting the moving image transmission.
2) When there is a retransmission request from the decoder 2 due to an error or the like, the picture is frozen on the decoder 2 side, and the encoder 1 side transmits a forced intra frame in response to the retransmission request. In this mode, the moving image transmission is resumed in the same manner as described above by shifting to the moving image transmission mode.

(F) Model-based prediction mode 1) Cutout of texture source The texture source is obtained from the analysis result of the moving area (in [A-3] extraction of target area) described later with respect to the first front image taken before the start of the call. Based on the image information of the selected person area.
2) Model selection and adaptation A 3D (three-dimensional) model is defined in advance by giving depth information to each 2D template 69a (front image).

3) Model-based prediction and transmission using 2D template The distance information can be calculated using a table calculated using camera characteristics for the front image using the template number and scaling value of the 2D template 69a. For this front image, a model-based predicted image of a human image can be generated by using distance information based on the region extraction result and 2D movement amount on the image plane. In addition, 2D model-based image transmission can be performed even when radio wave conditions are not good.

4) Model-based prediction using 3D motion information For example, the feature region of the 2D template 69a of the front image of a person is determined in advance as shown in FIG. 16, and the head and shoulders of the upper body template are determined based on the 2D motion information. 2 are regarded as rigid bodies, and a perspective transformation matrix representing a 3D position and orientation can be obtained from a set of 2D positions of N feature regions. The perspective transformation matrix can be obtained in this way by using a well-known technique such as a spatial quantization method.
Furthermore, in this case, such a thing may be stored in advance as a table, and such communication position / posture estimation is performed by controlling the robot in the factory. Therefore, it is based on the grounds that there is a premise that it is good if the user does not feel uncomfortable visually.

5) Background memory In a situation where the camera 3 is fixed and the background image that is normally captured is fixed, for example, when the camera 3 is fixed in a car, the target area priority mode is set. Background memory usage mode can be used.

6) Compensation of Forced Intra Frame Transmission Delay Time Based on Model As shown in FIG. 17, in the case of a car phone or the like, there is usually a session preparation time of 10 seconds or more before the start of a call. Therefore, using such session preparation time, the front face is transmitted in a forced intra frame (about 30 kbits) at the start of the session (not at the start of the call). At the same time, motion region analysis, template correction, 3D model adaptation, etc. are completed by the start of the call.

  Also, an approximate 3D motion vector based on the distance information L and 2D motion information is calculated. Thereby, the frame immediately before the start of the call is predicted on the model basis. Based on this, if the first frame is encoded in the video transmission mode after the start of the call, the amount of code can be reduced compared to when transmitting as a compulsory intra frame (at least in the background area is considered to be effective). .) Freezing, lowering of frame rate and image quality can be greatly alleviated.

(4) H. Compatibility with that of the H.261 standard In the encoder 1 in this embodiment, a bit stream of 64 kbps or less according to the protocol can be easily converted to H.264. For example, protocol converters 70 and 71 as shown in FIG. 18 can be configured to maintain compatibility.

  H. In the present situation, there is no uniform conversion means from the bit stream of the image signal of the H.261 standard to the protocol in the decoder 2 of the present embodiment. Can be connected.

  In the configuration of FIG. 18, for example, when the encoded bit stream transmitted from the encoder 1 is received, the protocol converter 70 adds a redundant code such as MBSTUFF to reduce the data amount of 9.6 kbps to H. . The H.261 standard is converted to a data amount of 64 kbps or more, or the variable length code syntax is changed to reduce the code amount. A simple conversion control that can be performed at a code level, such as replacement with one conforming to the H.261 standard, is performed. The signal is supplied to the H.261 encoder Sa.

  On the other hand, as described above, the protocol converter 71 does not have a uniform conversion means, but in order to transmit an intra picture, the image signal of the camera 3 is converted to the H.264 format. 261 decoder Sb. The output converted to the H.261 standard encoded bit stream is converted to conform to the transmission rate of 9.6 kbps by adding syntax change, mode information, etc. and reducing the frame rate to reduce the transmission code amount. Output.

  By doing so, H.C. It is possible to perform mutual conversion of image signals between the encoders Sa and decoders Sb of the H.261 standard and the encoders 1 and 2 of the present embodiment. Image communication can be performed in the same manner with a H.261 apparatus.

(5) State determination and mode transition (a) Forced setting For example, in the coding mode described above, mode switching between 1) texture source clipping mode and 2) model selection and adaptation mode is The target region extraction result can be automatically switched depending on whether or not the extraction result is a person, but this can also be forcibly switched by the user's manual setting.

(B) Determination of target area category In this determination operation, the mode progresses in consideration of each condition in accordance with the state transition diagram of the mode control related to the use environment and the subject as shown in FIG. .
1) Probabilistic state transition If there is a lack of judgment information necessary for state transition, autonomous state transitions using the transition probability table selected based on the probability given by default or incomplete information Perform a search. This is executed, for example, corresponding to the following case.
A) When there is no forced mode given from outside
B) When region extraction based on dynamic region analysis is not completed
C) When a signal analysis result such as a color signal is not completed 2) Deterministic state transition based on determination information As shown in FIG. A moving region analysis over about a frame time is performed as described later. After extracting a moving region and a movement occurrence state, a human region is extracted, and a template is selected by extracting shape parameters. The degree of conformity of the template is obtained by evaluation calculation, and the process ends when the standard is satisfied. If not, select the next template and repeat the same process.

(C) Update of transition probability value based on propagation of determination result based on state transition When a certain determination result causes a significant error in the subsequent evaluation calculation, it returns to the initial node of the moving object determination through the dynamic region analysis module. At this time, the transition probability (in the figure, “0.9” is set on the person side and “0.1” is set on the other object side as the branch of the moving object n) is updated. For example, if an area that is determined to be a person based on a purely probabilistic state transition in the initial determination outputs an evaluation function value that is unlikely to be a person due to subsequent 2D template matching, color analysis results, speed / acceleration evaluation, etc. The transition probability for determining “like” is changed and set to be lowered from “0.9” to about “0.2”.

(6) Human interface leading to mode selection Optimal coding control is possible even with the state determination and mode transition described above. However, the probability that an erroneous determination is made by the mode control unit 33 through guidance via the human interface 34 will be described. It can be drastically reduced. For example, if the following process is completed by the mode control unit 34 during the preparation time until the start of the session, the mode transition relating to the subsequent encoding control can be performed smoothly.

1) The user selects the person mode and sees the position direction of the camera 3.
2) A message “sit in front” is notified by voice or display via the human interface 34.
3) If the person is off the center of the screen, a message “Please turn slightly to the right / left” is notified via voice or display via the human interface 34. (As a display method, a child screen can be provided at the corner of the screen to indicate the position.)
4) Send the first frame as a forced intra.
5) A message “Connected to the other party. Please speak.” Is notified by voice or display via the human interface 34 (the call can be started).

(7) Transmission of mode information Mode information and accompanying template information can be transmitted because the amount of information does not interfere with ultra-low rate transmission. In this case, the transmission is performed in the user data area of the PIC hierarchy. Is transmitted in a state described in PSPARE (see FIGS. 30 and 50).

[A-3] Extraction of Target Area A target area to be watched, such as a person, is extracted based on a motion vector and camera settings. Here, in order to reduce the amount of calculation, the calculation is performed in units of macroblocks (MBK). The causal relationship between this calculation operation and various modes is as shown in FIG. 19 as described above.

(1) Region Extraction Based on Motion Vector Generation Here, it is a process in which the optimum template is finally determined according to the flowchart of the motion region extraction routine shown in FIG. Reference is also made to the area analysis explanatory diagram of FIG. 24 and the basic template diagram of FIG.

(A) Maximum motion vector detection and main region estimation In a coding standard such as H.261, the presence / absence of motion vector generation is included as information in MTP (macroblock type) as data indicating the attribute of the MBK. In this case, H.C. In the H.261 standard, data indicating that a motion vector is generated when the MTP value indicates values of 2, 3, 6, 8, and 10 (see FIG. 33). Therefore, H.H. As a motion vector detection result of H.261 encoding, for example, an arrangement of MBK attributes (MBA, MTP, CBP) as shown in FIGS. 20 to 22 are arranged in a row corresponding to each MBK position on the screen in which 12 rows of 6 rows of GOB are arranged and 33 rows of MBK of 11 rows and 3 rows are arranged in each GOB, respectively. , CBP value.

  Here, an area including a block having the largest motion vector is defined as a main area. In the evaluation of the magnitude of the motion vector, a panning vector that is a motion component accompanying the movement of the camera 3 is subtracted in advance. There are various known calculation methods for the panning vector. For example, the panning vector can be obtained by calculating an average of motion vectors at the peripheral portion of the screen as a simple calculation method.

(B) Analysis in time direction In the above-described case, it is difficult to specify the object region by examining the block attribute array for only one frame. Therefore, as shown in FIGS. 24A and 24B, analysis in the time direction is performed for about three frame times (see also FIG. 19, step D1 in FIG. 23). In particular, for the motion vector generation region, the position of the center of gravity is obtained (step D3) for the pattern (step D2 in FIG. 23) obtained by superimposing three frames, and the motion vector for each block position is as shown below. Is determined based on the number of determinations set in (step D4).
A) When a block whose distance from the center of gravity exceeds a certain threshold is 2 times or more in 3 times.

  Next, the isolated point area and the noise area are removed from the target area based on the obtained motion vector (step D5), the center of gravity of the remaining target area is calculated again (step D6), and the height of the target area is further increased. Find h and width w. Thereafter, for example, if it is a human head, the 2D template head is scaled by (h, w) to determine the approximate distance L0 to the camera 3 (step D7). Thereafter, in order to perform template fitting with higher accuracy, the following is performed (step D8).

(C) Collation and scaling of 2D template First, a block pattern composed of motion blocks is determined, and feature quantities are extracted. Here, if the person mode is manually selected in advance, the basic template of the person mode can be selected from the 2D pattern database 27 as shown in FIG.

  Here, as a basic template, a 2D template of a front image such as only a person's upper body, whole body, and head is prepared as a default pattern according to a mode predicted in advance by a decision tree (see FIG. 19). Next, the shape of the selected template is adaptively corrected. This correction includes center of gravity alignment, scaling, and local expansion / contraction (see FIG. 24C). Here, in the selection and correction of the optimum template, the value of the evaluation function F shown in Equation (2) described later is calculated.

A) Start from a template L (<L0) that is slightly smaller than the template obtained by the analysis in the time direction.
B) The template is overlaid on the image obtained by extracting only the motion blocks, and the number N _MBK (L) of motion vector blocks included in the region is calculated.
C) Increase the scale (decrease the distance L).
D) Proceed to the next step when L is below a certain value Lmin. Otherwise, return to a).
E) The point where the next evaluation function F takes the maximum value Fmax is taken as the optimum L.
F = −B × N _MBK (L) (2)
Here, B is a secondary differential value with respect to L of the curve of N _MBK (L).
Or, if you change this to a discrete expression,
F = −B (n) × N _MBK (Ln) (2a)
B (n) = A (n) -A (n-1)
A (n) = N _MBK (Ln) -N _MBK (Ln-1)
It becomes.

(D) 2D template including quantization scale information Although a 2D template can be expressed in binary, as described later, each MBK attribute array value can also be expressed in a template corresponding to a model in advance. As described later, for example, a quantization scale as shown in FIGS. 42 and 43 can be considered.

(E) Analysis of 2D movement pattern As a pattern in which a person moves from the screen, as shown in FIG. 25, horizontal horizontal movement (corresponding to the figure arranged in the horizontal direction in the figure), movement in the camera optical axis direction (Corresponding to the figures arranged in the vertical direction in the figure) can be considered.

(2) Generation of 2D template based on 3D model As shown in FIGS. 12 and 13, in the person mode under a limited environment, a 2D template equivalent to FIG. 25 can be created by perspective transformation of the 3D model. it can. This makes it possible to acquire parts such as the upper body, head, eyes, mouth, and nose. However, except for the front image, it is necessary to obtain the position and orientation of the target person with respect to the camera 3, but here the contents are not mentioned as being dependent on a general method.

(3) Determination of category of target region (a) Determination based on state transition graph In the state transition diagram of mode control shown in FIG. 19, a determination tree for specifying the category of the target region is included as a partial diagram. The transition between nodes on the determination tree is performed using two types of searches, the probability search based on the default probability and the probability search based on the determination information.

(B) Judgment using the MBK attribute Among blocks that have changed, the block value that has a particularly fine texture change is expected to have a high CBP score value. It can be used as a basis for determination of a partial area.

(4) Background Memory When a forced intra frame is transmitted, an area extraction process is always performed, and in particular, a process of storing background portion information in the background memory 32 is performed in the encoder 1. Such processing is also performed in the background memory 66 of the decoder 2. This is intended to substitute the data stored in the background memory 32 for the background information in the frame in order not to increase the amount of data transmission when it is determined that the movement or color change has become intense.

  That is, as shown in FIGS. 26, 27, and 28, all the background areas that are currently encoded are set to FIXD, and the use of the background memory is declared in the user data area. Based on this, the latest background information is called from the background memory 66 on the decoder 2 side, and the image is reproduced by overlaying it with target area information such as a person image that has been transmitted.

  FIG. 26 shows an outline of encoding control for achieving a very low rate, and a background region and a person region can be obtained by analyzing a motion region based on a plurality of frames of image signals and extracting the region. For example, the quantization control for each region is performed based on this, and the number of significant coefficients is controlled. On the other hand, encoding control is performed according to the buffer amount while setting the encoding rate and adding mode information. When extracting a person area, the area is specified in units of GOB MBK in the CIF format as shown in FIG. 27A, so that the background area can be obtained in the same manner (FIG. 27B). reference).

  FIG. 28 conceptually shows the case where the data stored in the background memory 32 is used, and the background area data obtained by performing the area analysis on the encoder 1 side is the background memory 32 (the background memory 1 in the figure). In the middle of transmitting the moving image after starting the moving image mode, new background area data is taken in and stored at any time. Similarly, the data in the background area is also stored on the decoder 2 side. When data transmission is interrupted during communication due to deterioration of the communication path, the background memory is stored on the decoder 2 side. The screen is reproduced with the data of the background area stored in the above, and a retransmission request is transmitted to the encoder 1 side to compensate for image reproduction until the moving image transmission is restored.

[A-4] Encoding control operation The encoding control operation is shown in FIG. FIG. It is a general thing which shows the syntax of the encoding bit stream in H.261 standard. In the following, details of the encoding control operation based on these FIGS. 29 and 30 will be described item by item.

(1) Change of bitstream syntax Redundant headers in the H.261 syntax (see FIG. 30) are deleted for ultra-low rate. For this reason, the GOB arrangement in the CIF structure is not changed, but for the 12 GOBs constituting the screen as shown in FIG. 31, the GBSC (16 bits) as the GOB start header and the GOB number code (4 The amount of data is reduced to half by adding the number of bits) only to the left GOB and eliminating the right half (see FIG. 32). As a result, 20 × 6 = 120 bits can be reduced per frame.

(2) Adaptive switching of variable length coding (VLC) table (a) Code word replacement The variable length code of the MTP (macroblock type) attribute of the H.261 standard is shown. For example, according to the inventor's experimental results at an extremely low rate, the probability of occurrence of each MTP value for a person image is in the order shown in FIG. Therefore, the number of bits for attribute description can be reduced by replacing the variable-length code for the MTP value according to the order shown in FIG. For example, by limiting the number of replacement patterns to a few, and writing which data to select in the user data area, the decoder 2 can be adapted to be changed depending on the situation. It can be done flexibly.

  In the experimental results, the MTP value is mostly 2 or 3 in the object region, and in the case of the human upper body mode, the total of both reaches about 100 to 150, so that the code length is 1 for each. And 2 (currently 2 and 3), the number of bits can be reduced by about 200 to 300 bits per frame.

(3) Adaptive control of attribute determination characteristics (a) Adaptive control of MTP determination using region information 1) For the background region, it may be set to FIX (fixed) or NOMC (no motion compensation) using region information. it can.
2) For the target region, MC-CO (motion compensation and coding) is determined for a person, particularly a face.

(B) Adaptive control of INTER / INTRA determination Usually, H.264 of 64 kbps or higher. FIG. 35 shows a characteristic curve for determining INTER / INTRA recommended in the H.261 standard. In this case, PVAR represents the inter-frame prediction error power sum for the four MBK Y blocks, and QVAR represents the value of the intra-frame variance of the input image for the four MBK Y blocks. Then, at a rate of 64 kbps or less, determination is made as follows.
1) Insert forced INTRA periodically to prevent error propagation.
2) The threshold value TH_INTRA is set as follows, for example. The filling rate RB of the FIFO buffer 11 is obtained by the following equation (3), and the threshold value TH_INTRA is set by determining the filling rate RB based on the comparison reference values RB1, RB2, and the like.
RB = BUFF / BUFF_MAX (3)

In this case, RB, BUFF, and BUFF_MAX indicate the current filling rate, the current number of used bits, and the maximum bit capacity for the FIFO buffer 11, and are determined according to the following three cases.
<Case 1> 0 ≦ RB <RB1
<Case 2> RB1 ≦ RB <RB2
<Case 3> RB2 ≦ RB
The threshold value TH_INTRA is set as follows corresponding to each case according to the determination result.
<Case 1> TH_INTRA = 64 × 256
<Case 2> TH_INTRA = 64 × 256 (Main part of the target area)
TH_INTRA = 64 × 512 (other than the main part of the target area)
TH_INTRA = 64 × 1024 (background area)
<Case 3> TH_INTRA = 64 × 256 (Main part of target area)
TH_INTRA = 64 × 1024 (other than the main part of the target area)
TH_INTRA = 64 × 2048 (background area)

In the above-described case, when the region information is not clear, as shown in the item of the image center priority mode or the target region centroid mode, a region having a certain distance from the centroid is defined as the target region. Then, the threshold value TH_INTRA is controlled as follows.
R_IR = IR_MBK / IRMBK_MEAN (4)
NTH = TH_INTRA × (R0 + R_IR) (5)
RPQ = PVAR / (QVAR × R_IR) (6)
Here, IR_MBK indicates the distance from the centroid of the target area to the MBK currently being encoded, and IRMBK_MEAN indicates the average distance from the centroid of the target area to all MBKs.

And for the value obtained from the above formula,
PVAR ≦ NTH or RPQ_VAR ≦ 1.00
Is determined to be INTER, otherwise it is determined to be INTRA. In this case, for example, the value of R0 is set to 0.5.

(C) Determination of presence / absence of motion compensation Usually, H.264 of 64 kbps or higher. The characteristic curve for motion determination recommended in the H.261 standard is as shown in FIG. In FIG. 36, the horizontal axis indicates the value of the error sum FDSUM between frames when there is no motion compensation, and the vertical axis indicates the value of the error sum MVSUM when motion compensation is performed. In this case, in the figure, the region with motion compensation (MC_ON) is set to a region surrounded by parameters GD1 and GD2 indicating the slopes of two straight lines and parameters IEV1 and IEV2 indicating threshold values for two FDSUMs. ing. At a rate of 64 kbps or less, each value of the parameter set (GD1, GD2, IEV1, IEV2) of the characteristic curve is changed as follows according to the buffer amount and the area. The buffer filling rate RB is expressed by the above-described equation (3).
RB = BUFF / BUFF_MAX (3)
In this case, RB, BUFF, and BUFF_MAX indicate the current filling rate, the current number of used bits, and the maximum bit capacity for the FIFO buffer 11, respectively.

<Case 1> 0 ≦ RB <RB1
In this case, conventional motion determination is used for all regions. Therefore, each value of the parameter set for the determination characteristic is set as follows.
(GD1, GD2, IEV1, IEV2) = (0.50, 0.91, 256, 256 x3)
<Case 2> RB1 ≦ RB <RB2
In this case, the normal motion determination characteristic is used inside the RM that is a slightly larger rectangular region surrounding the target region (automatically set for each template in consideration of the motion range between frames). Therefore, each value of the parameter set for the determination characteristic is set as follows.
(GD1, GD2, IEV1, IEV2) = (0.50, 0.91, 256, 256 x3)
In addition, in a region other than the above (corresponding to the background region) QM, the threshold for motion determination is set high.
(GD1, GD2, IEV1, IEV2) = (0.40, 0.80, 256x2, 256x4)
<Case 3> RB2 ≦ RB ≦ RB3
In this case, in the main area of RM,
(GD1, GD2, IEV1, IEV2) = (0.50, 0.91, 256, 256 x3)
Outside the main area of RM,
(GD1, GD2, IEV1, IEV2) = (0.40, 0.80, 256x2, 256x4)
In QM,
(GD1, GD2, IEV1, IEV2) = (0.20, 0.50, 256x4, 256x8)
<Case 4> RB3 <RB
In this case, one of the following is selected.
1) Switch to quasi-video mode and reduce frame rate
2) Switch to model-based mode
3) Set to picture freeze mode

(4) Use of past coding attributes (a) Characteristics of each coding attribute of the MBK layer When the MBK attribute of each frame stored in the attribute memory 22 is shown corresponding to the MBK position on the CIF, 22 per frame. The arrangement is x18.

1) MBA (macroblock address)
The MBA is assigned as a significant encoded block from the third frame. In particular, MBK (macroblock) in which MBA has a value of 2 or more often corresponds to the contour portion of the target area. Therefore, as shown in the flowchart of FIG. 37 and FIGS. 38 and 39, the MBA pattern one frame before is converted into a NOT_FIXED (NFX) pattern (step E1, FIG. 38, (a), (b) in FIG. 37). Reference), the average motion vector of the region information and the template information can be used to estimate and represent the MBA sequence one frame ahead (see steps E2 to E6 and FIG. 39).

  Then, encoding is controlled using such an estimated value of the MBA pattern, and the template information and the motion vector are encoded and transmitted (steps E7 and E8). On the other hand, on the decoder 2 side, the MBA array for one frame can be uniquely reproduced by the template information and the motion vector.

2) MTP (macroblock type)
A) INTER / INTRA mixed (normal frame)
Experiments have shown that at very low rates, most of the subject areas with motion (such as people) are labeled with MTP = 2,3. Therefore, based on the determination in the adaptive control according to the information generation amount described later, as shown in the flowchart of FIG. 40 and FIG. 41, the average motion vector and the template information of the region information are used unless particularly intense motion occurs. Thus, the MTP sequence one frame ahead can be expressed (see steps F1 to F3 and FIG. 41 in FIG. 40).

  Then, encoding for one frame is advanced based on the predicted MTP value (step F4), and the template information and the motion vector for each region are encoded and transmitted (step F5). On the other hand, on the decoder 2 side, the MTP arrangement for one frame can be uniquely reproduced by the template information and the motion vector.

B) Compulsory Intra Frame If a compulsory intra frame is declared as user data, then it is not necessary to describe data with MTP = 4 or data with MBA = 1. Thus, this makes it possible to save 5 × 22 × 18 = 1980 bits, ie about 2 kbits.

3) QSC (quantization scale)
The QSCs are all fixed to 31 (quantization step 62) below 32 kbps, and only the main region determined from the region analysis is set to 16 (quantization step 32). Therefore, QSC information is not transmitted even in a human mode of 16 kbps or less, even at an extremely low rate. Accordingly, it is not necessary to describe the QSC information in the GOB layer, and as a result, the information amount can be reduced by 12 × 5 = 60 bits. In addition, it is not necessary to transmit the change information of the quantization scale in MBK units.

  In this case, the region mode quantization in the person mode is shown as an example in FIG. 42A, and the quantization template is shown as an example in FIG. FIG. 43 shows an outline of data exchange between the encoder 1 and the decoder 2 when model-based transmission of a quantization template is performed. In the flowchart of FIG. 6 shows a flow routine relating to the setting of the value of the quantization scale QSC.

4) MVD (differential motion vector)
Since MVD is a horizontal motion vector change in the frame, a value other than “0” is generated in non-rigid body motion or rotational motion. For humans, movements caused by facial expression changes, three-dimensional rotation of the head and upper body edges correspond to this. Since these occur instantaneously, the prediction efficiency when predicting in the form of a difference vector is not very good. Therefore, such a difference vector is converted into a motion vector and predicted as an average 2D motion vector for each region.

  Here, the MVD generation position is limited to the MBK declared as having motion compensation in the MTP. Usually, most of this is represented by MTP = 2,3. As a result, the MVD information can be suppressed about once every two frames. FIG. 45 shows a flowchart of a routine for reducing the code amount by the average motion vector for each region, and FIG. 46 shows an outline of the contents in a pattern transition diagram.

5) CBP (encoded block pattern)
CBP indicates by bit information whether or not each of the 6 BLKs (blocks) included in the MBK is encoded. Therefore, by converting this CBP value into a YUV vector, it is possible to represent the CBP array one frame ahead using the average motion vector of the region information and the template information. On the other hand, on the decoder 2 side, the CBP arrangement for one frame can be uniquely reproduced by the template information and the motion vector. FIG. 47 shows a flowchart of the region-specific motion compensation prediction routine, and FIG. 48 shows the outline of the contents in a pattern transition diagram.

(B) Code amount reduction of MBK attribute using region information 1) Attribute encoding once in two frames As described above, by motion compensation using template information and average motion vector for each region, one frame worth Pattern attributes can be predicted and can be uniquely reproduced on the decoder 2 side. Note that all predicted pattern attributes can be reliably reflected in the encoding control not only on the decoder 2 but also on the encoder 1 side.

2) Alternate top and bottom prediction or interleaving Instead of putting the pattern attribute of prediction only once every two frames, as shown in FIG. 49 (a) (or FIG. 49 (b)) (Alternatively, the right side and the left side) can be alternately replaced with the prediction pattern. Although interleaving for each GOB line (see (c) to (e) in the figure) is also conceivable, there is a risk of discontinuity in the contour representing the pattern, so it is not adopted when the target area is large.

(5) Quantization control In principle, the quantization control is performed by controlling the quantization step. Here, the quantization step setting method is described in H.264. It is not stipulated in the H.261 standard, and its constraint condition is an even number of 2 to 62 (quantization scale value of 1 to 31) and a range that can be expressed by 5 bits. Therefore, in the present embodiment, the quantization control is performed by controlling the quantization step as described below.

(A) Region-specific quantization control 1) Target region priority mode In this mode, a small quantization step is assigned to the extracted target region. The quantization step is fixed at 62 for the background area. When the subject is in the person mode, only the head region is set to 62 or less, and the others are set to 62 in principle (see FIG. 42A).

2) Screen center priority mode In this mode, the closer to the screen center, the smaller the quantization step. However, in order to unify the control expression of the quantization step, a method of correcting the current buffer capacity for step calculation using the distance to the current MBK is adopted. Calculate using 7) and (8).

R_IR = IR_MBK / IRMBK_MEAN (7)
BUF_R = BUFF_MB
× (5.00 + real (IR_MBK) / real (IRMBK_MEAN)) (8)
here,
BUFF_MB: Buffer amount monitored in units of MBK BUFF_R: Virtual buffer amount based on distance calculation IR_MBK: Distance from target centroid to MBK currently being encoded IRMBK_MEAN: Average distance from target centroid to all MBKs This virtual modified buffer The value of the amount BUFF_MB is used in a control equation according to the encoding rate described later.

(B) Control according to the buffer amount Normally, the quantization control is performed in the above-described case, but the quantization control based on the buffer amount is not performed in the case of the forced intra frame transmission. Yes. The forced intra frame is normally transmitted in the following cases.

1) First picture when starting or resending moving picture mode 2) Semi-moving picture mode 3) Still picture mode (picture freeze)
4) Texture source image in model-based mode The quantization step depends on a control equation corresponding to the encoding rate described later.

(C) Control according to coding rate The determination formula of the quantization step (STEP FPIC) according to the coding rate (RATE) is set as follows.

1) Forced intra frame quantization step
1152kbps <RATE → STEP_FPIC = 12.0,
When 384kbps <RATE <1152kbps → STEP_FPIC = 14.0,
When 64kbps <RATE <384kbps → STEP_FPIC = 16.0,
When RATE <64kbps → STEP_FPIC = 32.0
2) Normal quantization step ISTEP = 2 × INT (BUFF_MB / (200.0 × QX64) +2 (9)
BUFF_MB: Current amount of data in the buffer QX64: Encoding rate = QX64 × 64.00 [kbps] satisfying the condition When the encoding rate is 16 kbps or less, changing the quantization scale frequently is the reverse This leads to an increase in the required number of bits. Therefore, the quantization step is fixed at 62 when the speed is 10 kbps or less.

(D) Adaptive control according to information generation amount Quantization and frame rate control are performed based on the amount of motion and the degree of color change.

1) Judgment of the intensity of motion The result of the determination based on this value by calculating the value of the index defined by the following equation (10) as the degree of the intensity of motion of the current frame relative to the past frame Quantization and frame control.

ただし、
Ｎｍｂ；動きの発生したブロックの数
Ｌ（Ｘ）；ベクトルＸのノルム関数．絶対距離，ユークリッド距離など
Ｖｉ；動きベクトル
Ｒｄ；伝送データレート
ＴＨＶ（Ｒｄ）；データレートに依存したしきい値定数
式（１０）で計算されるＡＭの値を用いて、新たに尺度ＡＭＴを計算する。この場合において、ＡＭＴは次のようにして計算される。

However,
Nmb: number of blocks in which motion has occurred L (X); norm function of vector X. Absolute distance, Euclidean distance, etc. Vi; motion vector Rd; transmission data rate THV (Rd); threshold constant depending on data rate A new scale AMT is calculated using the AM value calculated by equation (10). To do. In this case, AMT is calculated as follows.

A) When AM ≦ THV (Rd), AMT = 0
A) When AM> THV (Rd), AMT = AM
Here, the target range of Nmb and the corresponding THV are changed as follows according to the calculation capability of the encoding processor.

a) From the first MBK of the current frame to the MBK currently being decoded
b) From the first MBK in the current GOB to the MBK currently being decoded
c) All MBKs in the current GOB
d) All MBKs in the current frame
In the above cases a) and b), since a global operation is not required, the calculation amount is small and there is no processing delay, but the determination reliability is low. On the other hand, since c) and d) perform global calculations, the amount of calculation increases, but the processing delay is a maximum of one frame time. However, the reliability of judgment is high.

2) Determination of the intensity of the color change As a degree of the intensity of the color change of the current frame with respect to the past frame, the value of the index defined by the following equation (11) is obtained by calculation, and the determination is made based on this value. Based on the result, quantization and frame control are performed.

ただし、
Ｎｃｂ：ＣＢＰのブロック属性が１になったブロックの数
Ｃ（ｉ）：ｉ番目のマクロブロックに関してＤＣＴ係数のＤＣ成分の変化とＣＢＰに基づいてＹＵＶベクトルから色変化を計算する関数
ＴＨＣ（Ｒｄ）：データレートに依存したしきい値定数
式（１１）で計算されるＡＣの値を用いて、新たに尺度ＡＣＴを計算する。この場合において、ＡＣＴは次のようにして計算される。

However,
Ncb: Number of blocks whose block attribute of CBP is 1 C (i): Function for calculating color change from YUV vector based on change of DC component of DCT coefficient and CBP for i-th macroblock THC (Rd) : Threshold constant depending on data rate A new scale ACT is calculated using the value of AC calculated by equation (11). In this case, ACT is calculated as follows.

A) When AC ≦ THC (Rd), ACT = 0
A) When AC> THC (Rd), ACT = AC
Here, the target range of Ncb and the corresponding THC are changed as follows according to the calculation capability of the encoding processor.

a) From the first MBK of the current frame to the MBK currently being decoded
b) From the first MBK in the current GOB to the MBK currently being decoded
c) All MBKs in the current GOB
d) All MBKs in the current frame
In the above cases a) and b), since a global operation is not required, the calculation amount is small and there is no processing delay, but the determination reliability is low. On the other hand, since c) and d) perform global calculations, the amount of calculation increases, but the processing delay is a maximum of one frame time. However, the reliability of judgment is high.

3) Virtual buffer calculation
B) Virtual buffer increment based on motion
a) MBK of the target area without movement: BUF_M = 16 × (AMT / aM)
b) MBK of the target area with movement: BUF_M = 0
c) MBK of background area: BUF_M = 32 × (AMT / aM)
aM is a number corresponding to the average amount of motion per 1 MBK, for example, aM = 16.

B) Virtual buffer increment based on color variation
a) MBK of target area without color change: BUF_c = BMBK × (ACT / aC)
b) MBK of target area with color change: BUF_c = 0
c) MBK of background area: BUF_c = 2 × BMBK × (ACT / aC)
aC: A number corresponding to an average color change per MBK, for example, aC = 128
BMBK: Expected average code size per MBK, given by
BMBK = QX64 × 64000 / (Frate × NMBK)
Frate: Current frame rate
NMBK: Number of MBKs in one frame (6) Control of number of significant coefficients In 261, the DCT transform coefficients after quantization transform are zigzag scanned in units of blocks, and the obtained one-dimensional quantized coefficient sequence is a binary set (called an event) of a non-zero level followed by a length of zero run. expressing. Here, it is assumed that the coefficient of the high frequency component does not contribute much visually at an extremely low rate, and by limiting the number of events per block, the number of corresponding VLCs is reduced, and the number of bits is reduced as a whole. You can plan.

  That is, when the DCT significant coefficient (non-zero) obtained in order from the low frequency component by zigzag scanning exceeds a certain number, the remaining DCT coefficients are forcibly regarded as zero. At this time, control is performed so that the upper limit number Ncf (≦ 64) as the threshold value is switched according to each of the encoding rate, the region, the motion generation amount, the buffer amount, and the encoding mode. Note that the information regarding the upper limit number does not need to be sent to the decoder 2 side, and therefore is not encoded.

The above-described control of the number of significant DCT coefficients is actually performed as follows. Here, for example, the following state is assumed.
Encoding mode: People mode Encoding rate: 8 kbps
RB = V_BUFF / BUFF_MAX
V_BUFF = BUF_R + BUF_M + BUF_C
BUFF: Current buffer amount BUFF_MAX: Maximum capacity of buffer (RB1, RB2, RB3, RB4, RB5) = (0.2, 0.3, 0.5, 0.8, 1.0)
(Ncf0, Ncf1) = (16, 8)
In the determination, the following six cases are controlled according to the value of the buffer filling rate BF. The values indicated by RB1 to RB5 are determination thresholds, and values corresponding to the control contents are set in advance.

<Case 1> 0 ≦ RB <RB1
The maximum number of significant coefficients is 64 for all regions. <Case 2> RB1 ≦ RB <RB2
The maximum number of significant coefficients is 64 for the target area and the maximum is Ncf0 for the background area. <Case 3> RB2 ≦ RB <RB3
Maximum number of Ncf0 significant coefficients for all regions <Case 4> RB3 ≦ RB <RB4
Maximum number of Ncf1 significant coefficients for all regions <Case 5> RB4 ≦ RB <RB5
The background uses a background memory, and a portion not in the memory is expressed only by a DC component. Maximum number of Ncf1 significant coefficients for the target region <Case 6> RB5 <RB
Select one of the following << 1 >> to << 3 >> depending on other conditions, etc. << 1 >> Transition to quasi-video mode << 2 >> Switch to model base mode << 3 >> Perform picture freeze

(7) Frame rate adaptive switching (a) Description of frame rate change instruction In the bit stream syntax of the H.261 standard (see FIG. 30), an instruction to change the frame rate to the decoder 2 can be described by setting the TREF (temporal reference) value of the PIC layer. However, the frame rate change is treated as a subordinate means for reducing the ultra-low rate in this embodiment. The following shows the method and implementation factors.

(B) Method of changing frame rate After A / D conversion of a moving image on the encoder 1 side, a frame is obtained by thinning out by selecting whether or not to send raw image data to the encoding loop in units of frames. The rate is changed. Therefore, this thinning information is reflected in the TREF.

(C) Factors for driving frame rate change The drive factors for changing the frame rate can be summarized as follows.
1) Switching according to buffer capacity 2) Switching according to transmission rate (eg 8kbps → 5frame / sec)
In the moving image mode, an initial frame rate is set according to the transmission rate. For example, the following frame rate is set for the coding rate QX64.

・ QX64 ≧ 18 → 30 frame / sec
・ 18 ≧ QX64 ≧ 10 → 30 frame / sec or 15 frame / sec
・ 10> QX64 ≧ 6 → 15 frame / sec
・ 6> QX64 ≧ 1 → 15 frame / sec or 10 frame / sec
・ 64> QX64 × 64 ≧ 32 → 10-7frame / sec
・ 32> QX64 × 64 → 10 frames / sec or less 3) Switching according to the amount of motion generated 4) Mode change

[A-5] Description of Operation According to Encoding Process Program As described above, the functions of the ultra-low rate implemented in the encoding process are illustrated in the actual encoding process. The program is executed according to the flowchart of the program shown in FIGS. The outline of the overall flow will be described below.

  That is, first, the forced mode is set (step A1), and under this setting state, the target region extraction unit 26, the template database 27, the 2D template matching unit 28, the model base prediction unit 29, the three-dimensional shape database 30, and the like. Thus, analysis of the moving area and extraction of the moving area (steps A2 and A3) are performed. Next, the mode control unit 33 performs search and determination based on the state transition diagram, and subsequently sets the use environment mode, subject mode setting, state transition probability update processing, and encoding control mode determination (steps A4 to A4). A8) are sequentially executed.

  Thereafter, it is determined whether or not the current frame is a forced intra frame (step A9). If “YES”, the encoding control unit 21 determines the quantization step, and the orthogonal transform unit 7, The quantizing unit 8, the channel encoding unit 10, etc. 261 standard INTRA encoding processing is performed (steps A10 and A11). If “NO”, the process proceeds directly to the next virtual buffer calculation (step A12), and the encoding control unit 21 calculates the virtual buffer.

  Next, if the current frame is in the moving image mode, the frame rate is determined (steps A13 and A14). If the current frame is in the model base mode, the region extraction / recognition processing unit 25 determines the 2D template based on the 3D model. Generation is performed (steps A15 and A16), and 2D template matching is performed to extract the target region (steps A17 and A18). If the background memory mode is set, the background memory 32 is used (steps A19 and A20).

  Subsequently, it is determined whether or not the frame is an attribute prediction frame (step A21). If “NO”, a series of processes of steps A22 to A28 is executed. If “YES”, step A29 to step A29 are executed. A series of processing of A35 is executed. In steps A22 to A28, the encoding control unit 21 and the target region extraction unit 26 perform intra frame determination, motion compensation presence / absence determination, motion amount determination, color change amount determination, virtual buffer calculation, and quantization step calculation. Determination and calculation of the maximum number of coefficients are executed. In steps A29 to A35, the encoding control unit 21 and the attribute memory 22 and the like perform prediction of each sequence of MTP, MBA, QSC, MVD, and CBP, as well as encoding suppression processing of the MBK attribute, Model-based transmission is performed.

  After this, H.B. 261-based motion compensation and DCT calculation are performed (step A36), and region-specific quantization is performed by the orthogonal transform unit 7, the quantization unit 9, the inverse quantization unit 13, the inverse transform unit 14, the channel coding unit 10, and the like. After performing the control and the control of the number of significant coefficients, a bit stream of the BLK layer is generated (Steps A37, A38, A39). After the completion of the BLK layer, if it is an attribute prediction frame, MBK attribute determination control is executed (steps A40 to A42), and then MBK attribute data is stored and stored in the attribute memory 22 (step A43). ). Thereafter, encoded bit stream data for one frame is generated according to steps A44 to A54.

[B] Decoder Next, the content of the decoding process of the encoded bit stream received by the decoder 2 will be described. FIG. 4 conceptually shows the contents of the decoding process. When roughly classified, the decoding process is performed in three stages of a code level, a pattern level, and an image signal level, and various mode controls are performed in this process. In addition, the error correction function is implemented and the error correction function corresponding to the error rate at the radio communication level and the decoding process by the transmission at the ultra-low rate is realized. In the following, detailed explanations corresponding to the respective functions will be made item by item based on the conceptual configuration shown in FIG.

  The process of realizing the autonomous error correction function that is a feature of the decoder 2 will be briefly described. In other words, the average bit error rate is usually about 10 −6 to −8 in the wired transmission system, but on the other hand, it is assumed that the wireless system has a large value of about 10 −2 to −3. Yes.

  On the other hand, MPEG and H.264. In a bitstream generated with a syntax based on a variable-length code (see FIG. 30) as in the H.261 standard, even a single bit error propagates in the worst case, and subsequent data cannot be decoded. It is assumed that However, in a conventional wired video conference system, error correction to the extent that there is no practical problem has been realized by a combination of demand refresh (combination of freeze and retransmission request) and an error correction code (BHC code, etc.).

  In addition, error correction codes were sufficient for storage systems such as video CDs. However, in wireless systems, the error rate is high, and transmission errors such as omission, inversion, and insertion may occur in units of several bits depending on radio wave conditions. For this reason, complete error correction is difficult only with a normal code-theoretic approach. Therefore, in the decoder 2 of the present embodiment, a solution is achieved by taking an autonomous error correction means that can restore a signal value as close to the original signal as possible without interrupting the decoding process. It is.

  In the following, the contents of the autonomous error correction function will be described in the following items. [B-1] Global check of coded bit stream, [B-2] Error correction of code and syntax level, [B-3] Pattern level error correction, [B-4] Signal level error correction , [B-5] error correction strategy control using mode information, [B-6] recognition level error correction, [B-7] person memory and 3D model based method, [B-8] error correction strategy, [B-9] Operations according to the decryption processing program are divided into major items, and each principle and method are described by classifying the major items into smaller items.

[B-1] Global Check of Encoded Bitstream (1) Bit Error Occurrence Mode First, when correcting a bit error in a decoding process, which bit error of an image signal received via a communication path is determined The generation mode will be briefly described below.

(A) Single “0/1” inversion error: an error occurrence mode in which the bit value is inverted at random with a certain probability (b) Burst error: an error occurrence mode in which data in a certain section is masked The value between them is either << 1 >> or << 2 >> below. << 1 >> Outputs a fixed value of 0 or 1 continuously, << 2 >> Outputs a completely random value (c) Insertion / deletion of bits: This causes time expansion and contraction

(2) Basic policy of error correction In the present embodiment, the form of error generation is not particularly limited as the basic policy of error correction. However, the case (a) described above can be dealt with relatively stably, but in the cases (b) and (c), complete correction becomes extremely difficult. Is corrected with an estimated value which is not so strange visually, but error accumulation due to subsequent error propagation is expected, and means for suppressing this is provided. In this case, at the worst, a picture freeze and a retransmission request (ARQ: automatic repeat request) are exercised.

  Prior to exercising active correction means on the decoder 2 side, information having a large influence on the decoding process (data such as the required number of bits) is encoded by the encoder 1 as described above. By adding redundantly on the side, the correction capability can be improved.

(3) Transmission of the required number of bits Using the user data area in the protocol syntax, data that has a large effect on image reproduction when data is lost is redundantly transmitted from the encoder 1 side as described above. Yes. In the present embodiment, as shown in FIG. 50, the required number of bits in PIC units in the PIC layer and the required number of bits in GOB in the GOB layer are approximately 16 bits (up to 64 kbits) can be expressed in each user data area. Describe in. As a result, it can be determined whether or not there is any missing or inserted bit on the decoder 2 side. Whether or not the required bit number data itself has an error is determined as follows.

(A) As explained in the next section [B-2] Error correction of code and syntax level, the localization of PSC (picture start code) and GBSC (GOB start code) is determined by global header check in the buffer. Keep going. Based on this result, the code amount described in PSPARE or GSPARE is checked.
(B) If the sum of the code amount descriptions of the GOB layer does not match the code amount of the PIC layer, it is determined that there is an error in any description. If they match, it is determined that there is no error and the process ends.
(C) The average code amount per frame and the average code amount per GOB are calculated from the transmission rate and the frame rate. In a case other than a compulsory intra frame, a code amount that deviates from this value by a certain threshold or more is set as a description error location candidate.
(D) The candidate extracted in the above item (b) is compared with the linear prediction from the code amount over the past several frames, and if the deviation from the value is not within a certain range, an error is detected. judge.

(4) Forward type data check In order to detect all the data of the required number of bits in one frame described on the encoder 1 side before the start of the decoding process of one frame as described above (a) The bit stream data stored in the input buffer (FIFO buffer 42) on the decoder 2 side is globally checked in the order of several k bits. For that purpose, the localization of PSC and GBSC is required as described in the next item [B-2]. Therefore, the localization process is performed as follows.

    (A) A certain buffer area is completely scanned by fuzzy matching processing as shown in FIG. When the transmission rate is high, many matching calculations are required to complete localization by this method. However, since the amount of data per frame is several k bits or less at an extremely low rate, this way. Even if a global scanning process is performed, there is no practical problem.

  In this case, in the fuzzy matching process, the similarity S (see the formula in FIG. 51) indicated by the bit matching rate between the bit pattern of the code to be searched and the bit pattern in the encoded bit stream is 0.9 (90%), for example. ) If it is above, there is a standard for judging that there is almost no mistake. As a result, for example, since the PSC is 20 bits, it is possible to determine that the matching is allowed with an error of 2 bits. Note that the value of the similarity S is preferably 0.9, but in practice it can be implemented even when the value is set to a low value of about 0.8.

    (B) After searching for the PSC and the first GBSC, an approximate position is found based on the estimated code amount using the average code amount and the code amount one frame before. Thereafter, scanning is performed within a certain range centered on the position. If the number of GBSCs up to the PSC of the next frame is correct, the check process is terminated. If it is not correct, the value of the GN (group number) code is checked to determine the missing GBSC, scanning is performed from the GBSC immediately before that GBSC, and if it can be detected, the process ends.

[B-2] Level error correction of code and syntax (1) Localization of PIC and GBSC Only by decoding the bit stream sequentially, PIC data and There is a possibility that the range of the GOB data cannot be known. Therefore, a hierarchical process in which error correction based on the pattern level (MBK attribute) and image signal level (BLK) is pursued starting from the PSC and GOB headers that can be detected and localized relatively stably by the fuzzy matching process as described above. Is provided. After that, it is connected to the following error correction depending on the code and grammar.

(2) Generation form of grammatical decoding stop due to bit error When a bit stream including a bit error is sequentially decoded, the parser 43 stops the decoding process due to a grammatical error. However, the stop position at that time does not always match the bit error occurrence position, and in some cases, the bit error position may go back several tens of bits before the stop position. Therefore, in the following, what type of error occurrence exists will be described.

(A) When stopping immediately at a bit error position When a bit error is mixed in a fixed-length header and matching becomes impossible, decoding stops immediately unless there is a branch condition to others in the syntax.
(B) When stopping after decoding from a bit error position for a while 1) When a bit error occurs in a variable length code (VLC)

  As shown in FIG. 52, a bitstream is an output symbol obtained by transitioning between nodes based on logical conditions and numerical conditions in a graph obtained by combining code trees having different binary tree structures. It is a series. Therefore, if even one bit in the VLC is inverted, the time series of events (corresponding to encoding attributes and image signal values) that occurred on the encoder 1 side cannot be reproduced on the decoder 2 side, which is a completely different event. There is a risk that the sequence may be output as a decoding result.

  However, such an error cannot be detected only by a grammatical constraint that purely determines only a code. In other words, it is detected only with grammatical constraints via error detection conditions at the signal level or pattern level, which will be described later, or numerical conditions based on the image data format. In other words, since the erroneous bit stream decoding is continued until such an error detection process is started, the decoding stop position is behind the bit error position.

  In addition, even if such an error exists, the decoding is not necessarily stopped. For example, when the code is replaced with another codeword having the same code length even after bit inversion, the status output If only different values are output and the values do not adversely affect the subsequent conditional branching, no synchronization shift occurs in the subsequent VLC decoding process. Therefore, in this case, for example, even if there is a possibility that only the color or texture data of a certain pixel block is changed and reproduced, decoding does not stop.

2) When a bit error occurs in a fixed-length code In this case, the code length is the same as in the case of bit inversion, and the numerical value or attribute of the decoded output is different from that at the time of encoding. Decoding will not be stopped unless it adversely affects.

(3) Detection of grammatical decoding error Based on the protocol syntax of H.261 (see FIG. 30), the positions where bit errors occur are classified and explained as follows.

(A) Fixed length header that always appears 1) PSC (picture start code; 20 bits)
As long as the fuzzy matching process is performed, even if a bit error of about 2 bits occurs in the PSC, the PSC can be detected without depending on the syntax and the decoding result so far. Therefore, the detection and localization of PSC is an initial process necessary for detecting bit errors at other positions.
2) GBSC (group of blocks start code; 16 bits)
Similar to PSC, it can be detected stably by fuzzy matching processing, but if PSC localization is not performed, localization may be erroneous.

(B) Fixed-length data that always appears 1) TR (temporal reference; 5 bits)
If PSC localization is performed, it is easy to check the value because it is the following 5-bit data. The determination of error differs depending on the mode setting state on the decoder 2 side as follows.
A) In the normal moving image playback mode (fixed frame rate), it should be increased from the previous TR by a value corresponding to the frame rate corresponding to the transmission rate. If it is an ultra-low-rate moving image mode of 32 kbps or less, it becomes an increment of about 3 to 5, so it can be determined that the other is an error.
B) In the case of the semi-moving image mode, the increment is about 10 to 31.
2) PEI (picture extra insertion information; 1 bit)
If the data is 1, the following user data PSPAR (8 bits) exists. If the data is 0, GBSC follows.

3) GN (group number; 4 bits)
An error is determined in the localization process of GBSC. It is an error if the following conditions << 1 >> and << 2 >> are not satisfied. << 1 >> In the CIF structure, 1 ≦ GN ≦ 12, << 2 >> When one numerical value is increased from the previous GN 4) GQUANT (GOB layer quantizer infomation; 5 bits)
The quantization scale (QSC) in ultra-low rate transmission is fixed to 31 (quantization step 62) in the target region priority mode, and is set to 16 (quantization step 32) only in the main region determined from region analysis. To do. Therefore, QSC information is not transmitted in the person mode. This eliminates the need for GQUANT (GOB layer QSC information) and MQUANT (MBK unit quantization scale change information), so that no errors relating to this data occur. In other modes, when MQUANT is used, an error is determined by estimating a value in the same manner as CBP in pattern level error correction described later.

5) GEI (GOB extra insertion information; 1 bit)
A) If the data is “1”, then the user data GSPARE (8 bits) follows. Therefore, it is determined to be incorrect (synchronization error or error) from the definition of GBSC localization and the definition of GSPARE.
B) If the data is “0”, MBA continues. Therefore, the error determination is carried over to the MBK layer.

(C) Fixed length data / header that appears conditionally 1) PSPAR (picture layer spare infomation; 8 bits)
In the present embodiment, this area is used for mode information (8 bits) and required bit number information (8 bits × 2). As a result, an error in the required bit amount can be determined.

2) GSPARE (GOB layer spare infomation; 8 bits)
In an area where user data can be described in units of GOB layers, in this embodiment, the number of bits required in the GOB is described. Since the required number of bits in the GOB layer is likely to be within 8 bits, the bit pattern of the data of the required number of bits is described as an inverted pattern in which the order is changed between the upper (MSB) side and the lower (LSB) side. It is supposed to be. Therefore, the next GSPARE is required only when the required number of bits in the GOB layer exceeds 8 bits.

3) MBSTUFF (macroblock address stuffing; 11 bits)
Not used for very low rate transmission. H.64 kbps or higher. Although there is a possibility of being used in the H.261 standard, as shown in FIG. 53, even if a 1-bit error occurs, there may be a grammatically different interpretation result, so it is dangerous to perform fuzzy matching processing. . Therefore, the MBSTUFF code is not used in this embodiment.

4) MQUANT (MBK layer quantizer infomation; 5 bits)
As described in GQUANT, MQUANT is not used in the person mode for ultra-low rate transmission in this embodiment. When MQUANT is used in other modes, an error is determined by estimating a value in the same manner as CBP in pattern level error correction described later.

5) INTRA-DC (8 bits)
Grammatically, it appears as the first DCT coefficient data only when MTP indicates INTRA. This determination is left to a signal level error correction process described later.

6) ESC (6) + RUN (6) + LEVEL (8 bits)
If an error occurs in the ESC, another interpretation occurs in the subsequent decoding process of the BLK layer. Therefore, it is quite difficult to identify and correct the error position grammatically. Therefore, the following measures are taken.
A) In the DCT coefficient decoding process, the similarity Sesc with the ESC is always calculated. When Sesc = 5/6 (5 bits in 6 bits match the ESC), all the positions Pesc are stored, and the subsequent decoding is performed. Proceed with the process.
A) If the following conditions << 1 >> and << 2 >> are not satisfied before EOB is found, it is determined as an error, the process returns to the above Pesc, interprets it as ESC, and performs the decoding process again. << 1 >> Number of significant coefficients ≦ Ncoef constraint, << 2 >> Cumulative number of quantized DCT coefficients in BLK ≦ 64
C) If an error is detected in the image signal level error correction process (BLK layer), the process returns to Pesc, interprets it as ESC, and performs decoding again.
D) If an error is detected by the pattern level error correction process (MBK layer), the process returns to Pesc, interprets it as ESC, and performs decoding again.
7) EOB (end of block; 2 bits)
Since the word length is short, it is difficult to determine candidates based on similarity. However, since the number of appearances is large, the probability that a random error occurs in the EOB is not small. Therefore, it is determined that an error has occurred in the EOB when the following conditions << 1 >> to << 3 >> are not satisfied.

<< 1 >> Number of Significant Coefficients≤Ncoef Restriction << 2 >> Cumulative Number of Quantized DCT Coefficients in BLK≤64
<< 3 >> (number of BLKs in MBK) ≦ (number of BLKs described in CBP)
In this case, there are two correction methods (a) and (b) below. Usually, the method (b) is selected to simplify the calculation.
A) The bits are sequentially inverted immediately after the previous EOB, and the EOB pattern “10” is detected. Then, the detected EOB pattern is regarded as EOB and decryption processing is performed. If all the above three conditions << 1 >>, << 2 >>, and << 3 >> are satisfied, it is determined that this is a correct EOB.
B) Pattern level error correction is applied to all remaining MBKs in the GOB. When the above three conditions << 1 >>, << 2 >>, << 3 >> cannot be detected, it is left to signal level or pattern level error correction.

(D) Variable length code (VLC) that always appears
1) MBA (macroblock address; 1 to 11 bits)
Since MBA appears under the following conditions << 1 >> and << 2 >>, the VLC table is checked under this condition, and the rest is left to pattern level error correction.
<< 1 >> When the number of decoded MBKs is 32, immediately after EOB << 2 >> immediately after GEI = 0 2) MTP (macroblock type; 1 to 10 bits)
Since MTP appears under the condition “immediately after MBA”, the VLC table is checked under this condition, and the rest is left to pattern level error correction.

(E) Conditionally appearing variable length code (VLC)
1) MVD (motion vector data; 1 to 11 bits)
Since MVD appears under the following conditions << 1 >> and << 2 >>, the VLC table is checked under this condition, and the rest is left to pattern level error correction.
<< 1 >> When MTP = 2, 3, 6, 8, 9, 10 Immediately after MTP << 2 >> In MVD, two VLCs appear in succession in the order of x component and y component 2) CBP (coded block pattern; 3-9 bits)
Since CBP appears under the following conditions, the VLC table is checked under this condition, and the rest is left to pattern level error correction.

3) TCOEF (INTER) (transform coefficients; 2 to 17)
The VLC of the DCT coefficient for the INTER block of the BLK layer appears under the following conditions << 1 >> and << 2 >>. This error correction is left to signal level error correction.
<< 1 >> When MTP is 4 or 7 (INTRA block), the second and subsequent coefficients in the BLK layer << 2 >> When MTP is other than 4, 7, the next code is not ESC in the BLK layer [B-3 ] Pattern level error correction Use continuity between frames in hierarchical coding attributes.

(1) Use of Past MBK Decoding Attributes For past MBK layers that have already been decoded, five attribute data of MBA, MTP, QSC, MVD, and CBP are stored in the attribute memory 39c as array data in units of frames. ing. Since these attribute data have considerable continuity between frames, they have the property that they do not change so much except for scene changes and forced intra frames. This is especially true when there is temporal continuity regarding the existence of a target area such as a person area (see FIGS. 54 and 55).

  When an error of a signal level is detected in units of block pixels, if the cause of the error is not a VLC transmission error for a DCT coefficient, there is a high possibility that the error is an MBK attribute error. At this time, if MTP, QSC, and CBP are significantly different from each other in the range of the area composed of k MBKs adjacent to the motion compensation position in the MBK unit of the attribute one frame before, the past attribute is added to the attribute. Error correction is performed by applying a value. The comparison evaluation function in this case is as described below.

(A) Error detection and correction of MBA First, although MBA has temporal continuity in the left outline of the target area, it is expected that there will be many changes in the target area. However, on the contrary, the FIXED / NOT_FIXED pattern as described above can be almost predicted from the movement of the region. Therefore, as shown in the flowchart of FIG. 56 and FIGS. 57 and 58, in the following, NOT_FIXED (hereinafter referred to as NFX) uniquely corresponding to the MBA pattern by shifting the region using the average motion vector of the region. Pattern prediction calculation is performed, and the similarity SNFX between the NFX pattern based on the current decoding result and the NFX pattern based on the prediction result is calculated according to the following equations (12) and (13).

ここで、
L ；現在復号中のＭＢＡ値ａのＭＢＫアドレス（既に復号が完了した１つ前のＭＢＡ値のＭＢＫアドレスL0にａを加えた値）
Ls； NFXパターン比較の開始位置
ｓ(A,B) ； A=B のとき「１」，それ以外は「０」
NFXM(k) ；ＧＯＢの１番目のＭＢＫのＭＢＡ値が「１」以上ならば「１」，「０」なら「０」
NFXM＿(k) ；１フレーム前の NFXパターンから予測した NFXパターン
次に、上述の計算結果について信頼度ＲNFX を式（１４），（１５）にしたがって計算する。

here,
L: MBK address of MBA value a currently being decoded (value obtained by adding a to MBK address L0 of the previous MBA value that has already been decoded)
Ls: NFX pattern comparison start position s (A, B); “1” when A = B, “0” otherwise
NFXM (k); “1” if the MBA value of the first MBK of GOB is “1” or more, “0” if “0”
NFXM_ (k): NFX pattern predicted from the NFX pattern one frame before Next, the reliability RNFX is calculated according to the equations (14) and (15) for the above calculation result.

上述の結果に基づいて、例えば次のような４つの条件によりＭＢＡの検出と誤り訂正を行う。

Based on the above result, for example, MBA detection and error correction are performed under the following four conditions.

1) RNFX0 <0.5
In this case, it is determined that the reliability of the NFX prediction is low and is suspended. That is, for the time being, it is determined that the current NFX pattern is correct, and the process proceeds to determination of the next attribute.

2) RNFX0 ≧ 0.5 and SNFX <TNFX1
It is determined that the current NFX pattern is a decoding error. The NFX value is copied from the predicted pattern and converted to the MBA pattern. (TNFX1 is a value of about 0.3, for example)
3) RNFX0 ≧ 0.5 and TNFX1 ≦ SNFX <TNFX2
Since it cannot be determined that the current MBA value is a decoding error, it is suspended. That is, for the time being, it is determined that the NFX value is correct, and the process proceeds to determination of the next attribute. (TNFX2 is a value of about 0.7, for example)
4) RNFX0 ≧ 0.5 and TNFX2 ≦ SNFX
It is determined that the current NFX value is a correct decoding result.

(B) MTP error detection and correction At 10 frames / sec, the motion vector between frames is a three-frame vector at the video rate, so the maximum size is about 45 pixels. This corresponds to a maximum of 3 MBK displacements. Therefore, a comparison region composed of MBK that has been subjected to motion compensation with a motion amount (mx, my) obtained by converting the motion vector of the target region in MBK units with respect to the MTP pattern one frame before, and has already been decoded based on the MBK position. 59 and FIG. 60, and the corresponding region is compared with each MTP value by motion compensation. Here, the similarity SMTP is calculated based on the following equation (16), and further, the reliability evaluation value RMTP0 is calculated using equations (17) and (18) in order to evaluate the reliability of the previous MTP prediction. calculate.

here,
smtp (A, B): A function that calculates the similarity between two MTPs. The following score values are set and totaled for each of the six types of information included in the MTP.
INTRA attribute match → 3 points
Match of MQUANT attribute → 1 point
M V D Attribute match → 2 points
CBP attribute match → 2 points
TCOEFF attribute match → 1 point
F I L Attribute match → 1 point LMTP; Total value of score setting (here, “10”)
K: the number of MBKs included in the comparison area, the first is the MBK currently being decoded
K0: Number of MBKs included in the calculation region of reliability of MTP prediction and a value equal to or greater than K
MTP (i); MTP value of the first MBK in GOB, 0 in FIXED
MTP_ (i): MTP pattern predicted from the MTP pattern one frame before, usually the MTP pattern in the comparison region by motion compensation is used as the predicted pattern as it is. With the above setting, for example, between MTP = 1 and MTP = 2 Since the similarities of all match except for MVD and FIL,
3 + 1 + 0 + 2 + 1 + 0 = 7 (points)
It becomes. So in that case,
smtp (1,2) = smtp (2,1) = 7
It becomes. Similarly, other combinations can be calculated.

Next, based on the above result, MTP detection and error correction are performed under the following four conditions, for example.
1) RMTP0 <0.5
It is determined that the reliability of the MTP prediction is low and is suspended. That is, for the time being, it is determined that the current MTP value is correct, and the process proceeds to determination of the next attribute.
2) RMTP0 ≥ 0.5 and SMTP <TMTP1
<< 1 >> It is determined that the current MTP pattern MTP (L) is a decoding error. Copy the MTP value from the predicted pattern. (TNFX1 is a value of about 0.3, for example)
<< 2 >> On the other hand, all MTPs having a similarity of 8 or more (difference of 2 or less) are referred to, and the corresponding VLC is referred to.
<< 3 >> The VLC to be referred to and the bit stream string are compared by fuzzy matching processing, and the VLC having the highest matching degree is selected.
<< 4 >> If the matching degree satisfies a certain standard (0.8 or more), the VLC and MTP are adopted. If not, the first estimated MTP is used.

3) RMTP0 ≥ 0.5 and TMTP1 ≤ SMTP <TMTP2
Since it cannot be determined that the current MTP value is a decoding error, it is suspended. That is, for the time being, it is determined that the current MTP value is correct, and the process proceeds to determination of the next attribute. (TMBA2 is a value of about 0.7, for example)
4) RMTP0 ≥ 0.5 and TMTP2 ≤ SMTP
It is determined that the current MTP value is a correct decoding result.

(C) QSC error detection and correction All QSCs are fixed at 31 (quantization step 62) at a transmission rate of 32 kbps or less, and only the main region determined from region analysis in the target region priority mode 16 (quantization step 32) And Therefore, QSC information is not transmitted in the person mode. As a result, the QSC information in the GOB layer and the quantization scale change information in MBK units are not required, so that no QSC error occurs.

(D) Error detection and correction of MVD Since MVD is expressed by a differential motion vector between adjacent MBKs in a frame, it is difficult to determine an error with the data format as it is. Therefore, a method of evaluating the MVD data after returning it to the original motion vector MV is taken. Since MVD attribute data can be regarded as a signal value having statistical properties that are continuous in time and space compared to the values of other MBK attributes, linear prediction from a past frame and from surrounding MBKs can be considered. Linear interpolation is possible. Therefore, first, when there is no region information, the motion vector MV is calculated according to the following equation (19).

In the above formula,
vx (L, M); horizontal motion vector for the Lth MBK at M frame time vy (L, M); vertical motion vector for the Lth MBK at M frame time A; subscript representing x or y vA_ (L, M); Estimated value of motion vector for Lth MBK at M frame time a (i); Linear interpolation coefficient in frame b (m); Linear prediction coefficient between frames u; Intraframe interpolation and Interframe prediction ratio (0 ≦ u ≦ 1)
K: number of MBKs in surrounding MBK region including MBK being decoded p: past number of frames for performing linear prediction, and setting method of surrounding MBK are the same as those in the above-mentioned MTP comparison region. The value of the estimated vector thus obtained is evaluated using the error evaluation formula of the following formula (20).

式（２０）により得られた誤差評価の値Ｅにより次のように判定する。

A determination is made as follows based on the error evaluation value E obtained by the equation (20).

1) E ≧ 20
<< 1 >> It is determined that the decrypted MVD is an error, and vA is replaced with vA_.
<< 2 >> The MVD corresponding to this is calculated, and the corresponding VLC is referred to within a range of ± 5 with the center value as the MVD.
<< 3 >> The referenced VLC and the bit stream sequence are compared by fuzzy matching processing, and the VLC with the highest matching degree is selected.
<< 4 >> If the matching degree satisfies a certain standard (0.8 or more), the VLC and MVD are adopted. If not, the MVD of the first center value is adopted.

2) 20> E ≧ 10
Hold. For the time being, the decrypted MVD is held.
3) 10> E ≧ 0
It is determined that the decrypted MVD is correct.

  Next, when there is region information, the motion vector is calculated based on the following equation (21) instead of the equation (19), and the evaluation is performed in the same manner as described above.

ここで、
ｖRA(L,M-m) ；ｍフレーム前の時刻における領域の平均動きベクトル

here,
vRA (L, Mm); average motion vector of the region at time m frames before

(E) Error detection and correction of CBP 1) When there is no area information In the motion compensated predictive coding process, CBP data can be considered as a numerical value indicating the degree of temporal change in texture and color. However, since this CBP data does not have an algebraic structure that can perform linear interpolation calculation like MVD, as shown in FIGS. Perform the evaluation calculation. Therefore, when converting into a YUV vector, calculation is performed according to the following equation (22).

here,
c (L, M); YUV vector for Lth MBK at M frame time
c_ (L, M); YUV vector estimate for Lth MBK at M frame time
ac (i); Linear interpolation coefficient in the frame
bc (m); Linear prediction coefficient between frames
uc: Ratio between intra-frame interpolation and inter-frame prediction (0 ≦ uc ≦ 1)
Kc: number of MBKs in the surrounding MBK region including the MBK being decoded
pc: number of past frames for linear prediction
LN (i, m): A function for associating a number indicating the number of the i-th MBK in the comparison area at the M frame time and the address in the GOB, and if the comparison area is set, can be uniquely determined. The surrounding MBK can be set according to the MTP comparison area described above. The value of the estimated vector thus obtained is evaluated using the error evaluation formula of the following formula (23).

ｄ１(A,B) ；ベクトルＡとベクトルＢとの間の絶対値距離
式（２３）により得られた誤差評価の値Ｅにより次のように判定する。なお、ＹＵＶベクトルの定義から、
１２≧Ｅｃ≧０
であるので、以下のような判定を行う。

d1 (A, B): Absolute value distance between vector A and vector B Based on the error evaluation value E obtained by equation (23), the following determination is made. From the definition of YUV vector,
12 ≧ Ec ≧ 0
Therefore, the following determination is performed.

A) Ec ≧ 7
<< 1 >> It is determined that the decoded CBP is an error, and c_ is replaced with c.
<< 2 >> A CBP is calculated within a range of ± 1 with the replaced c as a center value (a plurality of CBPs may exist for one c), and the corresponding VLC is referred to.
<< 3 >> The referenced VLC and the bit stream string are compared by fuzzy matching processing, and the VLC with the highest matching degree is selected.
<< 4 >> If the matching degree satisfies a certain standard (0.8 or more), the VLC and CBP are adopted. If not, the first central value CBP is adopted.

B) 7> Ec ≧ 4
Hold. For the time being, the decrypted CBP is held.
C) 4> Ec ≧ 0
It is determined that the decrypted CBP is correct.

2) When there is region information Next, when there is region information, the calculation of the YUV vector is performed based on the following equation (24) instead of the equation (22), and the evaluation is described above. Do the same.

cR(L,M-m) ；ｍフレーム前の時刻における動き補償領域のＹＵＶベクトル
ただし、ｕｃの値は、領域情報がない場合の値よりも若干小さい値に設定することが望ましい。また、ｐｃは逆に少し大きく設定する。

cR (L, Mm); YUV vector of motion compensation region at time before m frames However, it is desirable to set the value of uc to a value slightly smaller than the value when there is no region information. On the contrary, pc is set a little larger.

(F) Processing of pending items As a result of detecting five attributes and evaluating errors as described above, for example, when three or more pending attributes occur, the attribute in which the pending appears. Only when the attribute value is replaced with the corrected error value, the VLC is referred to and evaluated again.

[B-4] Signal Level Error Correction The pattern level error correction described so far is error correction depending on the encoding attribute in MBK units, and is not a means for evaluating the value of the image signal itself. . The image signal can be obtained only when the bit stream data of the BLK layer is decoded at the code level and restored as a quantized DCT coefficient block. Therefore, error correction at the level of the image signal restored in this way will be described next.

(1) Error correction based on the properties of block pixels (a) Error detection Normally, decoding errors in the BLK layer are based on grammatical constraints and VLC tree code properties, and decoding errors in the DC component of the quantized DCT coefficient block Leads to As a result, even if MBK is grammatically decoded, its color (the DC component of the UV block is dominant) and texture are likely to be very different from the surroundings (see FIG. 66). ).

  On the other hand, as shown in FIG. 65 (b) as a normal image signal, a color different from that of the surrounding blocks is accidentally generated only in a tile-like pixel on the pixel block grid (see also FIG. 65 (a)). It is thought that there is very little probability. This way of thinking can be said that the reliability is still higher when the MTP of the target MBK indicates INTER. This is because there is a high possibility of being normally determined to be INTRA when the change in pixel value is MBK. Therefore, here, an error is detected as follows by paying attention to the color signal.

        1) As the most straightforward method, inverse quantization and inverse DCT are performed to restore the image signal block, and then the color is evaluated. This evaluation is based on a comparison between MBK estimated by linear interpolation from surrounding MBKs and motion compensated linear prediction from MBKs of past frames, and MBK restored from the bitstream at the present time. Specifically, error evaluation calculation is performed according to the following equations (25) to (29).

However, each symbol is as follows.
d1 (A, B); absolute value distance between vectors A and B Cb; color vector of restored MBK based on bitstream Ce; estimated MBK color vector Y $; average pixel value of Y block U $; Average pixel value of U block V $; Average pixel value of V block
YBLK $ (n, k); k-th pixel value of the n-th Y block in MBK (n = 1 to 4, k = 1 to 64)
UBLK $ (k); kth pixel value of U block in MBK
VBLK $ (k); k-th pixel value of V block in MBK For the error evaluation value Ec obtained as a result of the above calculation,
765 (= 255 × 3) ≧ Ec ≧ 0
In this range, for example, when Ec> 400, it is determined that there is an error. Note that the MBK prediction estimation calculation is performed in the same manner as in 2) below.

        2) A YUV vector is constructed using only the DC component of the quantized DC coefficient block, and an error between the linear prediction from the surrounding MBK and the YUV vector estimated by the motion compensation prediction from the past frame is expressed by the following equation (30). Therefore, calculate and evaluate.

However,
c (L, M); YUV vector for Lth MBK at M frame time
c (L, M) = (Y, U, V) where Y is the average of the DC components of the four BLKs,
U and V are DC components of UBLK and VBLK, respectively.
cR (L, Mm); if there is region information, it is the YUV vector of the motion compensation region at time m frames before, and if there is no region information, it is equivalent to c (L, M)
c _ (L, M); YUV vector estimate for the Lth MBK at M frame time
ac (i); Linear interpolation coefficient in the frame
bc (m); Linear prediction coefficient between frames
uc: Ratio between intra-frame interpolation and inter-frame prediction (0 ≦ uc ≦ 1)
Kc: number of MBKs in the surrounding MBK region including the MBK being decoded
pc: Past number of frames for performing linear prediction And the setting method of the surrounding MBK is in accordance with the method of setting the comparison area of the MBK attribute in the next item. The value of the estimated vector thus obtained is evaluated using the error evaluation formula of the following formula (31). The error determination conditions are the same as in 1).

(B) Error correction When it is determined that the decoded BLK is an error, it is replaced with an estimated value. On the other hand, in order to decode the next BLK, it is necessary to determine the bitstream decoding restart position at the code level. For this purpose, the EOB detection process in code level error correction is activated. At this time, if the number of reproduced MBKs (including corrections) is 22 or more in the GOB (the number of 2 columns in the 3 columns of blocks forming the GOB is 22), the remaining All MBKs are replaced with motion compensation prediction using region information.

(2) Prevention of error propagation due to periodic forced intra If decoding processing by frame addition is continued even after error correction as shown in the previous section, normal image reproduction becomes difficult due to accumulation of error propagation . Therefore, the following error propagation prevention measures are considered.
(A) By periodically inserting the forced intra MBK based on the area information, it is possible to prevent the image failure of the important part. In particular, as shown in FIG. 67, in the face portion, forced intra blocks are intensively assigned to the mouth and ears, and the increase in the amount of data is suppressed by cyclically circulating.
(B) After error correction is performed on the decoder 2 side, the encoder 1 side is requested to send the region-based forced intra in a distributed manner over a certain time range. In this case, since a significant increase in the amount of information can be avoided as compared with ARQ (request signal) in units of frames, there is no need to perform picture freeze.

[B-5] Strategic control of error correction using mode information As described above, the mode information set in the encoder 1 is divided into the three categories of usage environment, subject, and encoding control. By using the mode information, the error correction route can be specified on the decoder 2 side.

(1) Usage environment (a) Fixed in the car By setting the background memory 66 and the person mode, it is possible to activate error correction functions for all levels of recognition, pattern, and signal.
(B) In-car movable There is a possibility of both a portrait mode and a landscape mode. However, the data in the background memory 66 must always be updated and stored.
(C) Indoor fixed Background memory 66 can be used. Even when the subject is not a person, a template or a 3D model can be specially set depending on the purpose, so that error correction using area information similar to the person mode can be activated.
(D) Indoor movable The error correction function can be activated only in the person mode.

(2) Subject As described in the use environment described above, since the portrait mode is divided into the portrait mode and the landscape mode is difficult to specify the shape and color, the error correction function is activated in addition to relying on the background memory 66. difficult. In the person mode, features such as templates, 3D models, person image textures, eyes, nose, ears, hair, and skin color can be used, so that all levels of correction functions can be activated.

(3) Coding control mode Since it is not necessary to perform error correction in real time in a mode different from the normal real-time video playback request, for example, still image mode or video mail mode, it is sufficient to start ARQ. .

(4) Determination of the error occurrence state in the communication channel When the image cannot be corrected by other error correction means and decoding is stopped, the decoder 2 side is forcibly changed from the decoder 2 side to the encoder 1 side. One of the following request signals is transmitted. The transmission of such a request signal can be performed by automatically selecting and setting the level according to the condition of the communication channel, or can be set according to the user's preference. This is performed via the human interface 34.

(A) Retransmission request for moving picture transmission starting from forced intra picture This is the same as retransmission (ARQ) used in the existing scheme. However, ultra-low rate transmission leads to a significant increase in delay time and frequent picture freezes, so it is rarely used especially in real-time transmission of 32 kbps or less. On the other hand, this request is applied in the case of non-real time transmission such as the video mail mode.
(B) New restart of moving picture transmission starting from forced intra picture When the decoding process is stopped, a prediction image is generated from a past frame for the remaining MBK, and picture freeze is performed. Until the forced intra picture is normally sent (start position is confirmed by PSC header), the data in the input buffer is continuously discarded.
(C) Request for transmission of quasi-moving image only by forced intra picture.
(D) Motion parameter transmission request in model-based encoding.

[B-6] Recognition level error correction (1) Identification of target area The target area detection result and mode control information set on the encoder 1 side are described in the user data area (PSPARE, GSPARE). Therefore, this is detected by performing a decoding process on the decoder 2 side. Then, the template (2D information) is selected and transformed based on these pieces of information, and the region information extracted on the encoder 1 side is reproduced (see FIG. 68).

(A) Selection of basic template Since the same set of binary templates is provided in advance between the encoder 1 and the decoder 2, both are detected by detecting an identification number designating the template. Templates can be used.

(B) Modification of basic template 1) Centering Centering is performed with a 2D vector (mx, my) representing the horizontal and vertical shift amounts of the template center in pixel units.
2) Scaling The basic template is enlarged or reduced by the ratio r with the center obtained as described above as the origin.
3) Modification Partially modify the width, height, posture angle, etc., using the shape parameters unique to each basic template.

(C) Part identification When the category of the object is known from the mode information and the template selection information, it is possible to know which part of the template corresponds to what part of the object. For example, when a person's upper body template is selected in the person mode, as shown in FIG. 16 used in the description of the encoder 1, a fine image region corresponding to a person's head, face, mouth, eyes, nose, etc. Can be specified.

(2) Activation of Error Determination In this embodiment, the image correction process is described by specifying the cause of error as a bit error at the code level. Therefore, error detection itself is not performed at the recognition level, but at the signal level, code level, or pattern level. The error correction of the recognition level is to provide various properties reflected in the region information and the image of the target object based on the region information, particularly when the error is evaluated by them. Therefore, the activation of the error correction process at the recognition level is performed by another level module.

(3) Error correction using region information (a) 2D motion compensation by using region information If it is known in which region the MBK currently being decoded is included, as described in the explanation of pattern level error correction, Motion compensation can be performed using the motion vector of the MBK in the area where decoding has already been completed. For example, if it is determined that the MBK being decoded is included in the face area of the head, a motion compensated prediction MBK can be created using the average motion vector of the head.

(B) Correction using part information If an area is known, it is possible to correct not only movement but also color, brightness, and texture errors. For example, if the skin color is analyzed from face area information by the previous frame, errors can be determined and corrected by linear estimation values and error evaluation as described in the section of signal level and pattern level. It becomes like this.

(C) Representation of region information Region information is represented in the same manner as the template representation method described in the description on the encoder 1 side. For example, in the person mode, the subject creates templates in four layers of the upper body, the head, and the face based on the whole body image. In addition, the face has the mouth, eyes, nose, cheeks and forehead as main parts, and the relative position in the front face is set. The above-described skin color analysis is calculated by calculating an average color based on a portion excluding the mouth and eyes.

(D) Calculation of relative position in person area As described above, the basic template before the deformation is described as shown in FIG. Thereby, the positions of all parts can be expressed as two-dimensional coordinates in the basic template.

(4) Error correction by background memory If the area information becomes clear, the background area can be specified. Therefore, the background area image information is stored in the background memory 66 of the decoder 2. This makes it possible to perform error correction similar to the above even when a bit error occurs during the background region decoding process.

[B-7] Person Memory and 3D Model Base Method (1) Registration of Person Image The first intra picture can be labeled with a person name and stored in the person memory 67 as person image data. Here, the person identification name is assigned and registered as a label with the same meaning as the authentication number at the time of encoding by using the human interface 34. Also, the area information in the intra picture is stored as a template number and deformation information. Furthermore, the function can be further improved by storing the date and time of contact (call) with the person at the same time, or by storing the voice information with the same label as long as the memory capacity permits.

(2) Calling a person image The person image information stored by registering in the person memory 67 can be called by a user as a business card image at an appropriate timing based on the person identification name. This makes it possible to remember the face of the other party who made the call once, and for example, can be used as a means for authenticating the other party when making a call again. It can also be used as a texture image in model-based decoding mode when transmission errors become severe.

(3) Model-based decoding mode Texture mapping of the above human image to a 3D model of a person and adding 3D motion information to generate a human figure as an auxiliary image reproduction means when transmission errors become severe . This 3D motion information may be given so that the decoder 2 looks like a human motion appropriately. It is also possible to control using 3D motion information extracted on the encoder 1 side.

(4) Spatial limitation based on 3D model If 3D motion information is given from the encoder 1 side, it is possible to perform region prediction in cases other than the front image that could not be expressed only by the 2D template information described above. That is, as shown in FIG. 13, the 2D region can be identified on the image plane by the perspective projection of the 3D model.

[B-8] Error Correction Strategy Based on the error determination results at the above levels, the error correction functions that are finally exercised can be summarized as follows. Each level number is a value that predicts the degree of fidelity of the reproduced moving image to the original image.

<Level 0> 3D model base playback <Level 1> Picture freeze <Level 2> Model base estimation <Level 3> Linear estimation / motion compensation in GOB units based on region information <Level 4> Linear estimation in MBK units based on region information Motion Compensation <Level 5> Correction of Code Level [B-9] Description of Decoding Operation According to Flowchart Now, as described above, the actual decoding process for the error correction function performed in the decoding process In FIG. 8, each function is executed according to the flowcharts of the programs shown in FIGS. The outline of the overall flow will be described below.

  That is, first, a fuzzy matching process is performed by the fuzzy matching unit 44 of the channel decoding unit 35 on the bit stream data stored in the FIFO buffer 42 to search for a PSC (step B1), and the parser is based on the search position. The required number of bits, mode information, and area information are decrypted by 43 (steps B2 to B4). Then, the global check is performed again based on these data, and the GBSC in one frame is localized (step B5). Thereafter, the determination operation for the code level error is performed by the error comprehensive determination routine shown in FIG. 11 (step B6).

  Next, the use environment, the object (subject), and the decoding mode are set by the mode control unit 33 based on the obtained mode information (steps B7 to B9). In the person mode, the basic template is selected, corrected, deformed, and scaled to reproduce the region pattern (steps B10 to B13). Thereafter, the pattern level error is passed through steps B14 to B17. Implement the correction function.

  First, the attribute estimation unit 61 detects and corrects MBA and MTP errors (steps B18 and B19). If the person mode is set, the error determination / correction unit 50 performs QSC suppression. Then, MVD error detection and correction are performed (steps B20 to B22), and then the CBP error detection and correction are performed by the attribute estimation unit 61 (step B23). Based on these results, the error determination / correction unit 50 performs comprehensive error determination according to the routine shown in FIG. 11 (step B24), and subsequently, when an attribute error is detected in any of the above-described processes. Return to that step again to execute error detection (steps B25 to B28), and when these are completed, the data of the decoded attribute array is stored in the pattern attribute unit 39c which is an attribute memory.

  Thereafter, after steps B30 to B33, the decryption processing unit 36 performs H.264 processing based on the data of the attribute array described above. The BLK decoding process according to the H.261 standard is performed (step B34). Next, when the person mode is set, the pattern / signal level error correction unit 37 estimates the texture and color based on the part information, the feature amount, and the linearity of the color vector based on the surrounding and past MBK. Estimation and color vector evaluation are performed (steps B36 to B38).

  If an error has occurred and LBLK is 22 or more, motion compensation estimation is performed for the remaining MBK in the GOB, and thereafter, the GOB ends (steps B40 and B41), and an error occurs. However, if LBLK is 22 or less or no error has occurred, ESC detection and correction and EOB detection and correction are performed, and then a comprehensive error determination routine shown in FIG. B42 to B44), the decoding process is continued by repeating this.

  According to this embodiment, when the decoder 2 side receives an image signal of a moving image via a digital communication path with a high transmission error rate and performs a decoding process, the existing moving image Compression standard An error correction function is provided at the code, grammar, pattern, signal, and recognition levels based on the protocol at the stage of decoding processing so as to receive an encoded bitstream with improved ultra-low rate in H.261. In addition, the pattern check, signal, and grammar error detection are driven by the global check of the encoded bitstream in consideration of the required bit amount, and the region information (person, etc.) based on the mode information and the 2D template is driven. It can be used to perform error correction based on recognition level evaluation.

  Also, according to the present embodiment, on the encoder 1 side, syntax change and code word replacement, prediction of encoding attribute of current frame based on past encoding attribute and adaptive control of attribute determination, motion And low-speed image transmission can be realized by comprehensively using object area extraction and area-specific quantization control by model and control of significant number of transform coefficients according to use mode, transmission rate, and motion generation amount. It will be like that. The encoder 1 according to the present embodiment is of a level that can be realized only by making a slight change to the current image compression standard (H.261). Therefore, a protocol converter is provided for the implementation. This can be realized with a simple configuration.

The block block diagram of the encoder which shows one Example of this invention Block diagram of decoder Conceptual diagram of the encoding process Conceptual diagram of the decryption process Flowchart of encoding processing program (part 1) Flowchart of encoding processing program (part 2) Flowchart of encoding processing program (part 3) Flow chart of decryption processing program (part 1) Flow chart of decryption processing program (part 2) Flow chart of decryption processing program (part 3) Flow chart of comprehensive error determination routine Explanatory drawing when the camera is fixed in the car as a usage environment Illustration of the three-dimensional positional relationship of people in the car Examples of camera images inside a car, indoors, and outdoors that include people Illustration of the effect of setting the area by the distance scale Illustration of person front view template and feature area Explanatory diagram when compensating for transmission delay in model-based mode H. Explanatory drawing of the structure for performing communication between H.261 terminals State transition diagram of mode control related to usage environment and subject H. Example of macroblock attribute array based on H.261 syntax (part 1) H. Example of macroblock attribute array based on H.261 syntax (part 2) H. Example of macroblock attribute array based on H.261 syntax (part 3) Flowchart of routine for moving area extraction and template determination Action explanatory diagram of moving area extraction and template determination Illustration of basic template and its deformation Illustration of the principle of ultra-low rate based on human area extraction (A) GOB, MBK lattice in CIF format and (b) Example of background memory image Illustration of the effect of using and updating the background memory Chart for explaining the concept of overall coding control H. H.261 standard bitstream syntax Explanatory drawing of the addition part of GOB number and header (GBSC) Comparison explanatory diagram with and without GOB header reduction H. Comparison table of variable length code (VLC) of MTP in H.261 standard Correspondence table between generation probability and code length of each MTP value for human image INTER / INTRA judgment characteristics Motion block judgment characteristic diagram Flow chart of MBA prediction routine Explanatory diagram showing the correspondence between MBA and NFX Explanatory drawing of prediction of MBA pattern from previous frame Routine flowchart for reducing MTP information Explanatory drawing of MTP code amount reduction by average motion vector for each region Illustration of quantization by region and quantization template in portrait mode Illustration of model-based transmission of quantization template QSC setting routine flowchart Routine flowchart for reducing MVD information Explanatory drawing of MVD code amount reduction by average motion vector according to area CBP prediction and evaluation routine flowchart Explanatory diagram of motion compensated prediction by region of CBP Explanatory drawing when performing interleaving for MBK attribute prediction Explanatory diagram of transmission form of required number of bits and global check Explanatory diagram of fuzzy matching process of PSC Explanatory drawing of combining binary trees with syntax and generating bitstreams Example of misinterpretation caused by bit error when MBSTUFF is used Explanatory diagram of continuity between frames in hierarchical coding attribute (encoder side) Explanatory diagram of continuity between frames in hierarchical coding attributes (decoder side) Flowchart of MBA error detection and correction routine Explanatory diagram showing the correspondence between MBA and NFX Explanatory drawing of prediction of MBA pattern from previous frame Explanatory drawing of evaluation of decryption result based on similarity calculation of MTP Explanatory drawing showing an example of scan order in evaluation calculation Flowchart of CBP error detection and correction routine Explanatory diagram showing the definition of CBP Explanatory drawing which shows the conversion process from CBP value to YUV vector Illustration of YUV vector prediction (A) GOB and MBK lattices in CIF format and (b) an example of block-like error pattern Explanatory drawing showing interpretation error of bitstream caused by bit error at image signal level and influence on signal error Explanatory drawing which shows the example of the periodic dispersion | distribution type | mold forced INTRA classified by area Flow chart of routine for area reproduction or pseudo area setting

Explanation of symbols

  1 is an encoder (encoder), 2 is a decoder (decoder), 3 is a camera, 5 is an A / D converter, 7 is an orthogonal transformer, 8 is a subtractor, and 9 is a quantizer. 10 is a channel encoder, 11 is a FIFO buffer, 12 is a channel, 13 is an inverse quantizer, 14 is an inverse transformer, 15 is an adder, 16 is a prediction memory, 17 is a loop filter, and 20 is a motion. Detection unit, 21 is an encoding control unit, 22 is an attribute memory, 23 is an attribute prediction unit, 24 is an encoding processing unit, 25 is a region extraction / recognition processing unit, 26 is a target region extraction unit, 27 is a template database, 28 Is a 2D template matching unit, 29 is a model-based prediction unit, 30 is a three-dimensional shape database, 31 is a person memory (person information storage unit), 32 is a background memory (background information storage unit), 33 is a mode control unit, and 34 is Human interface 35, a channel decoding unit, 36 a decoding processing unit, 37 a pattern / signal level error correction unit, 38 a recognition level processing unit, 39 a memory (storage means), 39a a mode information unit, 39b is an area information part, 39c is a pattern attribute part, 39d is a 2D motion vector part, 39e is a personal identification information part, 41 is a D / A converter, 42 is a FIFO buffer, 43 is a parser, 44 is a fuzzy matching part, 45 Is a storage unit, 46 is a comparison table, 47 is an error determination unit, 48 is an inverse quantization unit, 49 is an inverse transform unit, 50 is an error determination / correction unit, 52 is an adder, 53 is a prediction memory, and 54 is motion compensation. , 55 is a loop filter, 57 is a frame memory, 58 is a pixel value estimation unit, 59 is an image estimation unit, 60 is a motion vector estimation unit, 61 is an attribute estimation unit, 62 is a decoding control unit, and 63 is model-based prediction. Part 64 is a 3D shape database, 64a is 3D shape data, 65 is a person image database, 66 is a background memory (background information storage means), 67 is a person memory (person information storage means), 68 is an area reproduction unit, and 69 is a template. The database 69a is a 2D template.

Claims (22)

In the process of encoding the image signal, in the image recognition device that analyzes the image signal and recognizes an object existing in the region in the frame,
A mode condition for setting an incidental situation with a predetermined code is stored, and mode control means for recognizing the image signal under the condition is provided.
The mode control means includes
It is provided in advance to be able to set 3D model data that shows the target shape as a 3D rough model.
An apparatus for recognizing an image, characterized in that it is possible to set a model base mode in which position information and color information for the three-dimensional model data are generated as encoded data. The mode control means includes
The distance information is calculated based on a two-dimensional template corresponding to a screen set in advance in generating the target model-based predicted image in the model-based mode. Image recognition device. The mode control means includes
In the model base mode, by previously setting the feature region of the two-dimensional template, and roughly estimating the position and orientation of the three-dimensional model from the position data of the feature region based on the two-dimensional motion information, The image recognition apparatus according to claim 1, wherein the apparatus is configured to perform model-based prediction. The mode control means includes
The transition from the set mode state to another mode state is determined by the transition probability value set in advance according to the level when the information necessary for the determination according to the analysis state of the image signal is insufficient. 4. The image recognition apparatus according to claim 1, wherein the image recognition apparatus is configured as described above. The mode control means includes
When the transition determination between modes using the transition probability value is executed, if the error value becomes large by performing an evaluation calculation on the determination result, the transition probability value is updated and set so as to reduce the error. The image recognition apparatus according to claim 4, wherein: The mode control means includes
6. The image recognition apparatus according to claim 5, wherein the evaluation calculation for the determination result is performed by a two-dimensional template, a color analysis result, a speed / acceleration evaluation, or the like. A human interface capable of exchanging information with users has been established.
7. The mode control means according to claim 1, wherein the mode control means is configured to give setting information to a user via the human interface so as to be in an incidental situation corresponding to a predetermined mode setting state. The image recognition apparatus according to any one of the above. Target area extraction means for extracting a target area existing in the screen;
8. An attribute data generating means for reducing the amount of information at the time of encoding processing by setting attribute data of the target area extracted by the target area extracting means. An image recognition apparatus according to claim 1. The target area extracting means includes
The image recognition apparatus according to claim 8, wherein the target area is configured to be performed based on whether or not a motion vector is generated. The target area extracting means includes
The image recognition apparatus according to claim 9, wherein the target area is extracted by detecting a motion vector generation area over three frames or more. The target area extracting means includes
The image recognition according to claim 9 or 10, wherein the target region is detected as a block in which a maximum motion vector is generated, and is estimated as a main region including a block in which the maximum motion vector is generated. apparatus. The target area extracting means includes
Panning vector calculation means for calculating a panning vector indicating the amount of motion of the entire screen based on the average value of the motion vectors at the periphery of the screen,
12. The image recognition apparatus according to claim 11, wherein when detecting the maximum motion vector, the detection is performed after subtracting a panning vector calculated by the panning vector calculation means. The target area extracting means includes
It is configured to extract a block pattern composed of motion blocks, calculate a feature amount thereof, and extract the target region by scaling a corresponding template when there is template information. The image recognition apparatus according to claim 8. The target area extracting means includes
14. The image recognition apparatus according to claim 13, wherein the template information is configured to perform model-based transmission of data indicating block-unit attribute information corresponding to a model. The target area extracting means includes
The image recognition according to claim 14, wherein when extracting the target area, the movement is easily detected by analyzing a pattern in which the target in the screen moves two-dimensionally. apparatus. The target area extracting means includes
The image recognition apparatus according to claim 15, wherein a two-dimensional template is generated based on a three-dimensional model when extracting the target region. The target area extracting means includes
17. The image recognition apparatus according to claim 8, wherein the image recognition apparatus is configured to determine a category of the target area based on a control state according to the mode information. The target area extracting means includes
18. The image recognition apparatus according to claim 17, wherein the image recognition apparatus is configured to determine a category of the target area based on a state transition condition in mode control. The target area extracting means includes
19. The image recognition apparatus according to claim 17 or 18, wherein a category of the target area is determined based on attribute information set in units of blocks. The target area extracting means includes
9. The system according to claim 8, wherein region extraction processing is always performed when a forced intra frame is transmitted, and information on a background region estimated from the target region is designated and transmitted at that time. 20. The image recognition device according to any one of items 19 to 19. The target area extracting means includes
21. The image recognition apparatus according to claim 20, wherein the background area information is designated so that transmission of the background area is selectively performed according to the transmission amount of the information. The target area extracting means includes
22. The image recognition apparatus according to claim 21, wherein when the amount of information transmission is limited, data for combining the information designated as the background area information with the person information being transmitted is transmitted.

Images