JP2019004300A - Data number determination apparatus, data number determination method, and data number determination program - Google Patents

Abstract

To determine the number of data referred to when data is compressed.SOLUTION: A data number determination apparatus includes a separation processing unit that separates each of a plurality of pieces of past data referred to for predicting prediction target data to be predicted included in the time series data into a plurality of pieces of data, a prediction unit that generates prediction data for each combination of the number of pieces of past data used for processing and data after separation, a difference calculation unit that performs processing of calculating a difference between prediction target data having a time correlation with the past data and prediction data for at least each of the pieces of data after the separation, for each candidate number of the past data used for the processing, and a data number determination unit that determines the number of pieces of past data on the basis of the difference obtained for each of the number candidates.SELECTED DRAWING: Figure 6

Description

  The present invention relates to a data number determination device, a data number determination method, and a data number determination program.

  Acquisition of various and enormous data at each sampling time has become possible due to advances in network technology and sensing technology. One type of data that can be acquired at each sampling time is multidimensional time-series data (hereinafter referred to as “spatial information data”) including position information in space. The spatial information data is, for example, time-series data including the position information of the mobile terminal as the sensing device and the statistical information of the user of the mobile terminal, the position information of the vehicle of the car navigation system as the sensing device and the user of the vehicle Time-series data including statistical information, time-series data including location information of a ticket gate through which a user of a traffic IC card has passed and statistical information of a user of a traffic IC card.

  Spatial information data is expected to be used for marketing, infrastructure development, urban development and disaster support. When used in these fields, the meaning of the spatial information data is examined by analyzing the spatial information data with high accuracy according to the usage form. The data analyzer acquires and accumulates a large amount of spatial information data in order to analyze the spatial information data with high accuracy.

  An increase in the amount of acquired spatial information data leads to an increase in cost necessary for storing the spatial information data. For this reason, the accumulated spatial information data needs to be efficiently compressed. A data compression device, which is a device for compressing data, can use the correlation in the time direction in the spatial information data in order to efficiently compress the accumulated spatial information data.

  Spatial information data may be represented by an image. For example, the spatial information data may be expressed as an image representing a map in which bar graphs are distributed. The data compression apparatus models the correlation in the time direction of the spatial information data based on the audio or image coding standard. The data compression apparatus compresses the spatial information data using the modeled correlation in the time direction (see Non-Patent Document 1).

“H.265 / HEVC textbook”, Impress Japan, pp. 23

  However, the temporal correlation used in the audio or image frame compression technique may not apply to the temporal correlation of spatial information data. For this reason, the data compression apparatus cannot simply compress the audio or image frame data simply by diverting the compression technique for encoding the audio or image to the compression of the spatial information data. Also, the data compression device cannot take the lead against future events.

  In order for the data compression apparatus to efficiently compress the data of the audio or image frame and to take the lead against future events, the number of frames that the data compression apparatus refers to when compressing the data. Need to be determined. In other words, the conventional apparatus needs to determine the number of data referred to by the data compression apparatus. However, the conventional apparatus has a problem that the number of data to be referred to at the time of data compression cannot be determined.

  In view of the above circumstances, an object of the present invention is to provide a data number determination device, a data number determination method, and a data number determination program capable of determining the number of data referred to when data is compressed. Yes.

  One aspect of the present invention is a separation processing unit that separates each of a plurality of past data into a plurality of data, which is referred to in order to predict the prediction target data that is included in the time-series data. A prediction unit that generates prediction data for each combination of the number of past data used for processing and separated data, prediction target data having a temporal correlation with the past data, and at least a difference between the prediction data Based on the difference calculated for each candidate of the number of past data used for the processing, and the difference obtained for each of the number of candidates, the number of the past data is calculated for each subsequent data. And a data number determining unit that determines the number of data.

  One aspect of the present invention is the above-described data number determination device, wherein the data number determination unit determines the number of the past data that minimizes the sum of the differences between the separated data.

  One aspect of the present invention is the data number determination device described above, further including a separation position determination unit that determines a position at which the past data is separated, wherein the separation position determination unit is a candidate for a position to be separated. A separation position at which the sum of differences between the separated signals obtained when the data is separated is minimized.

  One aspect of the present invention is a data number determination method executed by a data number determination device that determines the number of data, in order to predict prediction target data, which is a target to be included, included in time-series data. A separation processing step for separating each of a plurality of past data to be referred to, a prediction step for generating prediction data for each combination of the number of past data used in the processing and the separated data, and the past A difference calculation step of performing, for each candidate of the number of past data used in the process, a process of calculating a difference between the prediction target data having a temporal correlation with the data and each of the separated data at least after separation. A data number determination method for determining the number of the past data based on the difference obtained for each of the number candidates. That.

  One embodiment of the present invention is a computer that separates each of a plurality of past data into a plurality of pieces of data, which is referred to in order to predict prediction target data included in time-series data. The difference between the prediction data, the prediction step for generating prediction data for each combination of the step, the number of past data used for processing, and the separated data, the prediction target data having a temporal correlation with the past data, and the prediction data Is calculated for each candidate of the number of past data used in the process, and the past is calculated based on the difference obtained for each of the number of candidates. A data number determination program for executing a data number determination step for determining the number of data.

  According to the present invention, it is possible to determine the number of data to be referred to when compressing data.

It is a figure which shows the example of a structure of the forecast apparatus in 1st Embodiment. It is a figure which shows the example of the prediction process and the forecast process in 1st Embodiment. It is a figure which shows the example of the virtual forecast process in 1st Embodiment. It is a figure which shows the example of the frame group containing a reference frame in 1st Embodiment. It is a flowchart which shows the example of operation | movement of the forecast apparatus in 1st Embodiment. It is a figure which shows the example of a structure of the forecast apparatus in 2nd Embodiment. It is a flowchart which shows the example of operation | movement of the forecast apparatus in 2nd Embodiment. It is a flowchart which shows the example of operation | movement of the forecast apparatus in 2nd Embodiment. It is a figure which shows the example of a structure of the forecast apparatus in 3rd Embodiment. It is a figure which shows the example of the forecast process based on a class in 3rd Embodiment. It is a figure which shows the example of the forecast process based on a class in 3rd Embodiment. It is a flowchart which shows the example of operation | movement of the forecast apparatus in 3rd Embodiment. It is a flowchart which shows the example of operation | movement of the forecast apparatus in 3rd Embodiment.

Embodiments of the present invention will be described in detail with reference to the drawings.
(First embodiment)
FIG. 1 is a diagram illustrating an example of the configuration of the forecasting device 1. The forecasting device 1 is an information processing device that predicts future data based on past data constituting time-series data. The forecasting device 1 notifies forecast signals representing future data to other devices such as data compression devices.

  The forecasting apparatus 1 includes an input signal storage unit 10, an order update unit 11, a prediction coefficient calculation unit 12, a virtual forecast signal generation unit 13, a virtual forecast error calculation unit 14, and a virtual forecast error minimization determination unit 15. The virtual forecast error minimum value storage unit 16, the optimum order storage unit 17, the prediction coefficient determination unit 18, and the forecast signal generation unit 19 are provided.

  The forecasting apparatus 1 includes an order update unit 11, a prediction coefficient calculation unit 12, a virtual forecast signal generation unit 13, a virtual forecast error calculation unit 14, a virtual forecast error minimization determination unit 15, and a virtual forecast error minimum value storage. The unit 16 is provided as the data number determination device 2 which is a device for determining the number of data. The prediction device 1 includes at least one of the input signal storage unit 10, the optimal order storage unit 17, the prediction coefficient determination unit 18, and the prediction signal generation unit 19 as a function unit of the data number determination device 2. Also good. The number of data is, for example, the number of audio or image frames. In the following, the number of data is the number of frames of an image as an example.

  Part or all of the functional units of the forecasting device 1 is realized by a processor such as a CPU (Central Processing Unit) executing a program stored in the storage unit. The storage unit includes, for example, a nonvolatile recording medium (non-temporary recording medium) such as a magnetic hard disk device or a semiconductor storage device. Some or all of the functional units may be realized by using hardware such as LSI (Large Scale Integration) or ASIC (Application Specific Integrated Circuit).

  The input signal storage unit 10 stores a plurality of frames as time series data. For example, the input signal storage unit 10 stores a plurality of frames as spatial information data. Each frame is divided into N blocks for each frame. The shape of the block is a rectangle. The data number determination device 2 executes a prediction process and a forecast process for each block.

FIG. 2 is a diagram illustrating an example of prediction processing and prediction processing. In FIG. 2, the frame at the current time among the plurality of frames 100 is a frame 100-t. Hereinafter, a target frame for which a pixel value is predicted is referred to as a “prediction target frame”. The time interval between the frames 100 is a unit time length such as 1 minute, for example.
[Prediction processing]
The prediction process is a process in which the data number determination device 2 generates the pixel value of the prediction target frame for each block based on the prediction target frame and the frame 100 that is earlier than the prediction target frame. In FIG. 2, the data number determination device 2 generates the pixel value of the frame 100-t for each block based on the reference result in the frame 100-t and the past frame 100- (t-1) or the like. To do.

  The order update unit 11 shown in FIG. 1 updates the forecast order p, which is the order used for the prediction process and the forecast process, in the order of p = 1,. P is a predetermined order. The prediction coefficient calculation unit 12 reads the frame 100-(t−1) immediately before the frame 100 -t from the input signal storage unit 10. The prediction coefficient calculation unit 12 reads one or more frames 100 determined by the prediction order p updated by the order update unit 11 from the input signal storage unit 10 as a reference frame.

  The prediction coefficient calculation unit 12 performs a prediction process for each block of the frame 100- (t−1) immediately before the frame 100-t based on the linear model shown in Expression (1). In the following, in order to simplify the notation, the frame data is expressed as a one-dimensional signal.

  Here, f (x, t) represents the pixel value of the pixel at the position x in the frame 100-t. x represents a value in the range of 0 ≦ x ≦ X−1. X represents a predetermined value for the pixel value. t represents a value in the range of 0 ≦ t ≦ T−1. T represents a predetermined value for time. P in Formula (1) represents the prediction order which is the order used for the prediction process. The prediction order p represents the number of frames 100 referred to in the prediction process. The predicted order is equal to the predicted order.

In the following, the symbol “^” written on the character in the mathematical expression is written immediately before the character. ^ F _p (x, t) represents a frame 100-t that is a prediction target frame. a ₀ (t), a ₁ (t),..., a _p (t) respectively represent prediction coefficients that are coefficients used in the prediction process. In the _{following, a 0 (t), a} 1 (t), ..., a p a _(t), referred to as _a p (t). The i-th block in the frame 100 satisfies Expression (2).

The prediction coefficient determination unit 18 illustrated in FIG. 1 calculates a p-th order prediction coefficient that minimizes a prediction error when the frame 100-t is a prediction target frame. Similarly, the prediction coefficient calculation unit 12 calculates a p-th order prediction coefficient that minimizes a prediction error when the frame 100- (t−1) is a prediction target frame. Minimizing the prediction error means, for example, setting the prediction error to a threshold value or less. Prediction coefficient calculation unit 12, the case of a prediction target frame frame 100- (t-1), so as to minimize the prediction error E _p represented by the equation (3), defining a prediction coefficient.

Prediction coefficient determining unit 18, based on the order p prediction parameters that minimize the prediction error E _p in the case where the prediction target frame frame 100-t, the equation (4), (5) and (6) To calculate. Similarly, the prediction coefficient calculation unit 12, a p-th order prediction parameters that minimize the prediction error _{E p} in the case where the prediction target frame frame 100- (t-1), formula (4), (5) And it calculates based on Formula (6).

Note that the prediction coefficient a _p (t) obtained as a solution of Expression (4) represents a p-th order prediction coefficient that minimizes a prediction error when the frame 100-t is a prediction target frame. For example, when p = 1, Expression (4) is expressed as Expression (7).

When p = 1, the prediction coefficient a _p (t) is expressed as Equation (8) as a solution of Equation (7).

  Here, Δ is expressed as in Expression (9).

[Forecast processing]
The forecast process is based on the past frame before the prediction target frame, based on the future frame after the prediction target frame, and the pixel value of the prediction target frame that is the future frame after the prediction target frame. Is generated for each block. In FIG. 2, the data number determination device 2 does not depend on the future frame 100-(t + 1), but based on the pixel values of the past frame 100 before the frame 100 -t, A pixel value is generated for each block.

The prediction signal generator 19 shown in FIG. 1 uses the prediction signal representing the result of the prediction processing of the frame 100- (t + 1), which is the prediction target frame in the prediction processing, as a solution of the equation (4). It produces | generates like Formula (10) based on _ap (t). Hereinafter, the prediction processing for the frame 100-t immediately before the prediction target frame is referred to as “virtual prediction processing”. The virtual forecast signal generation unit 13 similarly generates a virtual forecast signal that represents the result of the virtual forecast process for the frame 100-t.

  The forecast signal generator 19 calculates the forecast error of the frame 100- (t + 1) as shown in Expression (11). Similarly, the virtual forecast error calculation unit 14 calculates a virtual forecast error of the frame 100-t.

  FIG. 3 is a diagram illustrating an example of the virtual forecast process. In FIG. 3, the data number determination device 2 generates the pixel value of the frame 100-t for each block without referring to the pixel value of the frame 100-t within the frame.

  The virtual forecast signal generation unit 13 reads one or more reference frames determined by the forecast order p and the prediction coefficient calculated by the prediction coefficient calculation unit 12. The virtual forecast signal generation unit 13 executes a virtual forecast process for the frame 100-t based on the prediction coefficient and the reference frame. The virtual forecast signal generation unit 13 generates a virtual forecast signal representing the result of executing the virtual forecast process for the frame 100-t.

[Adaptive selection of forecast order]
A method for setting the forecast order p, which is the order used in the forecast process of Equation (10), will be described. The prediction coefficient calculation unit 12 acquires the forecast order p from the order update unit 11. The prediction coefficient calculation unit 12 calculates a p-order prediction coefficient a _p (t−1) that minimizes a prediction error when the frame 100- (t−1) is a prediction target frame.

  The virtual forecast signal generation unit 13 sets a prediction coefficient that minimizes a prediction error in the frame 100-t immediately before the prediction target frame as a prediction coefficient used for the prediction process of the frame 100- (t + 1). Note that the prediction error in the prediction process decreases as the prediction order increases in the prediction process. However, the relationship between the forecast error in the forecast process and the forecast order p in the forecast process is not clear.

  The virtual forecast signal generation unit 13 executes the forecast process for the frame 100-t immediately before the forecast target frame, instead of executing the forecast process for the frame 100- (t + 1) that is the forecast target frame in the forecast process. That is, the virtual forecast signal generation unit 13 performs a virtual forecast process for the frame 100-t. The virtual forecast signal generation unit 13 generates a virtual forecast signal representing the result of executing the virtual forecast process for the frame 100-t.

The virtual forecast error calculation unit 14 calculates a forecast error (hereinafter referred to as “virtual forecast error”) between the frame 100-t obtained from the input signal storage unit 10 and the virtual forecast signal representing the frame 100-t. The virtual forecast error calculation unit 14 calculates a p-th order virtual forecast error when the frame 100-t is set as a forecast target frame in the virtual forecast process based on the p-th order prediction coefficient a _p (t−1). .

The virtual forecast error minimizing determination unit 15 compares the calculated virtual forecast error with the provisional minimum value of the virtual forecast error. The virtual forecast error minimizing determination unit 15 determines the forecast order p ^* that minimizes the virtual forecast error as shown in Expression (12) by comparing the magnitude of the virtual forecast error for each forecast order p.

When the calculated virtual forecast error is smaller than the provisional minimum value of the virtual forecast error, the virtual forecast error minimization determination unit 15 determines the virtual forecast error that is smaller than the provisional minimum value of the virtual forecast error as the provisional minimum of the virtual forecast error. The value is recorded in the virtual forecast error minimum value storage unit 16 as a value. The virtual forecast error minimizing determination unit 15 records the forecast order p in the optimum order storage unit 17 when the calculated virtual forecast error is smaller than the provisional minimum value of the virtual forecast error. In this way, the virtual forecast error minimizing determination unit 15 estimates the optimal forecast order p ^* based on the virtual forecast error of the frame 100-t.

The virtual forecast error minimum value storage unit 16 stores a provisional minimum value of the virtual forecast error. The optimum order storage unit 17 stores the optimum forecast order p ^* . The forecast signal generation unit 19 executes the forecast process for the frame 100- (t + 1) based on the forecast order p ^* determined by the formula (12) and the formula (10).

[Use of periodic correlation]
The virtual forecast signal generation unit 13 may refer to each frame 100 in a cycle including the current time in the virtual forecast process. The virtual forecast signal generation unit 13 may refer to each frame 100 (non-neighboring frame) in the cycle before the cycle including the current time in the virtual forecast process.

  The virtual forecast signal generation unit 13 refers to the non-neighboring frame in the virtual forecast process based on the periodicity of the time series data. Prediction processing that refers to each frame 100 and a frame in the vicinity of each frame 100 from one cycle before the current time of the frame 100-t to M cycles before is represented as Equation (13).

Here, D represents a cycle such as one day, one week, one month, one year, or the like. Based on the equation (13), the virtual forecast signal generation unit 13 refers to (2p _m +1) frames in the vicinity including the frame 100- (t−M × D) that is M cycles before the frame 100-t. . p _m satisfies the expression (14).

  The virtual forecast signal generation unit 13 sets a prediction coefficient that minimizes a prediction error in the frame 100-t immediately before the prediction target frame as a prediction coefficient used for the prediction process of the frame 100- (t + 1). The virtual prediction error calculation unit 14 calculates the prediction error in the frame 100-t as shown in Expression (15).

Here, ƒ _p (x, t, a _p (t)) is expressed as in Expression (16).

[Adaptive selection of forecast order (use of periodic correlation)]
A method for setting the forecast order p, which is the order used in the forecast process based on the equation (13), will be described. 1 updates the forecast order p, which is the order used for the forecast process, in the order of p = 1,... The prediction coefficient calculation unit 12 calculates a p-order prediction coefficient a _p (t−1) that minimizes a prediction error when the frame 100- (t−1) is a prediction target frame. The virtual forecast signal generation unit 13 executes a virtual forecast process for the frame 100-t.

The virtual forecast error calculation unit 14 calculates a _p -th order virtual forecast error in the case where the frame 100-t is a prediction target frame based on the p-th order prediction coefficient a _p (t−1). The virtual forecast error minimizing determination unit 15 compares the magnitude of the virtual forecast error for each forecast order p to determine the forecast order p ^* that minimizes the virtual forecast error as shown in Equation (17).

The prediction coefficient determination unit 18 sets the prediction order p ^* stored in the optimum order storage unit 17 as the prediction order used in the prediction process of the frame 100- (t + 1) that is the prediction target frame. Prediction coefficient determining unit 18, based on the reference frame determined as the frame 100-t by forecast orders ^{p *,} so as to minimize the prediction error _{E p} of represented by the equation (3), the frame 100-t The prediction coefficient is calculated.

The forecast signal generation unit 19 generates a forecast signal of the frame 100- (t + 1), which is the forecast target frame, based on one or more reference frames determined by the forecast order p ^* and the prediction coefficient. The forecast signal generation unit 19 executes forecast processing as shown in formula (13) based on the forecast order p ^* determined by formula (17).

  FIG. 4 is a diagram illustrating an example of a frame group including a reference frame. The time-series data includes a plurality of frames 100 having a periodic time correlation with each other. The frame group 200 including the reference frame includes, for example, the frames 100 that are continuous with the frame 100- (t + 1) on the same day. The frame group 200 on the same day may have a periodic time correlation with each other.

  The frame group 201 including the reference frame may be configured of, for example, each frame 100 at the same time as the time of the frame 100- (t + 1) on a different day. Frame groups 201 at the same time on different days may have periodic time correlations with each other. The frame group 202 including the reference frame may be configured by, for example, each frame 100 having a time different from the time of the frame 100- (t + 1) on different days. Frame groups 202 at different times on different days may have periodic time correlations with each other.

  The forecast signal generation unit 19 is based on at least one of past data in a cycle including the time of the frame 100- (t + 1) as future data and past data in a cycle not including the time of the frame 100- (t + 1). Thus, frame 100- (t + 1) is generated. In FIG. 4, the cycle including the time of frame 100- (t + 1) is “today”. In FIG. 4, the period that does not include the time of frame 100-(t + 1) is “previous day” or “previous day”.

  Hereinafter, the block including the pixel value multiplied by the prediction coefficient on the right side of the equation (13) is referred to as a “forecast reference block”. Therefore, the forecast reference block is a unit block in the reference frame in the forecast process. Hereinafter, the block including the pixel value multiplied by the prediction coefficient on the right side of Expression (16) is referred to as “prediction reference block”. Therefore, the prediction reference block is a unit block in the reference frame in the prediction process.

Virtual Warning signal generator 13, to limit the combination of forecasts reference block in the process of determining the forecast orders p ^*, may reduce the amount of calculation processing for determining a forecast orders p ^*. For example, for the frame 100-t that is the prediction target frame in the virtual forecast process, the virtual forecast signal generation unit 13 calculates the similarity between the virtual forecast block that is a unit block in the frame 100-t and the prediction reference block. The virtual forecast signal generation unit 13 sorts the prediction reference blocks in descending order of similarity. That is, the virtual forecast signal generation unit 13 rearranges the prediction reference blocks in descending order of similarity. The order updating unit 11 can change the prediction order p with priority on the prediction reference block having a high degree of similarity, thereby limiting the reference blocks for each prediction order p to one set. The degree of similarity of the prediction reference block is expressed as, for example, Expression (18).

As described above, the prediction coefficient calculation unit 12 and the virtual forecast signal generation unit 13 refer to past data that is referred to when generating prediction data based on the similarity between the unit block in the prediction target data and the unit block in the past data. Limit the number of unit blocks in. Thereby, the prediction coefficient calculation unit 12 and the virtual forecast signal generation unit 13 can reduce the amount of calculation in the process of determining the forecast order p ^* .

[Exception handling]
When the pixel values of all the pixels in the prediction block are zero values as in Expression (19), the virtual forecast signal generation unit 13 does not execute the reference block sorting process (first exception process).

  The virtual forecast signal generation unit 13 sets the reference block selection order in a predetermined order when the pixel values of all the pixels in the prediction block are zero values. Further, the prediction coefficient determination unit 18 determines a prediction coefficient at a predetermined coefficient (second exception process). For example, the prediction order p of the frame 100- (t + 1) may be the same as the prediction order determined for the frame 100-t that is a unit time before the time of the frame 100- (t + 1). The virtual forecast signal generation unit 13 may set the weight of each reference frame to an equal value.

  When the similarity between the prediction block and the prediction reference block is less than the threshold, the virtual forecast signal generation unit 13 sets the prediction coefficient of the prediction reference block whose similarity with the prediction block is less than the threshold to zero (third) Exception handling). The threshold value is, for example, the same value as the minimum similarity.

Next, an example of the operation of the forecasting apparatus 1 will be described.
FIG. 5 is a flowchart showing an example of the operation of the forecasting device 1. The input signal storage unit 10 reads a plurality of frames 100 as time series data (step S101). The order update unit 11 updates the predicted order p in the order of predicted order candidates p = 1,..., P (step S102). By executing step S102 to step S109, the data number determination device 2 determines the forecast order p.

  The prediction coefficient calculation unit 12 reads a frame 100- (t−1) that is a frame immediately before the frame 100-t. The prediction coefficient calculation unit 12 reads one or more frames 100 determined by the prediction order p updated by the order update unit 11 as a reference frame. The prediction coefficient calculation unit 12 calculates a p-th order prediction coefficient that minimizes a prediction error when the frame 100- (t-1) is a prediction target frame (step S103).

  The virtual forecast signal generation unit 13 reads one or more reference frames determined by the forecast order p and the prediction coefficient calculated by the prediction coefficient calculation unit 12. The virtual forecast signal generation unit 13 generates a virtual forecast signal representing the frame 100-t as a result of executing the virtual forecast process for the frame 100-t based on the prediction coefficient and the reference frame (step S104).

  The virtual forecast error calculation unit 14 calculates a virtual forecast error that is a forecast error between the frame 100-t obtained from the input signal storage unit 10 and the virtual forecast signal representing the frame 100-t (step S105). The virtual forecast error minimization determination unit 15 compares the calculated virtual forecast error with the provisional minimum value of the virtual forecast error (step S106). If the virtual forecast error is greater than or equal to the provisional minimum value of the virtual forecast error (step S106: NO), the virtual forecast error minimization determination unit 15 advances the process to step S109.

  When the virtual forecast error is smaller than the provisional minimum value of the virtual forecast error (step S106: YES), the virtual forecast error minimization determination unit 15 sets the smaller virtual forecast error as the provisional minimum value of the virtual forecast error (step S106). S107). The optimum order storage unit 17 stores the forecast order p associated with the provisional minimum value of the virtual forecast error as the optimum forecast order p (step S108).

The prediction coefficient determination unit 18 sets the prediction order p ^* stored in the optimum order storage unit 17 as the prediction order used in the prediction process for the frame 100- (t + 1) that is the prediction target frame (step S110). The prediction coefficient determination unit 18 calculates the prediction coefficient of the frame 100-t based on the frame 100-t and the reference frame determined by the prediction order p ^* (Step S111). The forecast signal generation unit 19 generates a forecast signal of the frame 100- (t + 1), which is the forecast target frame, based on one or more reference frames determined by the forecast order p ^* and the prediction coefficient (step S112).

As described above, time-series data does not include future data. The time series data includes past data. The time series data further includes prediction target data in addition to the past data. That is, the time series data includes the prediction target data in distinction from the past data. In the embodiment, the past data and the prediction target data are continuous on the time axis, and the prediction target data and the future data are continuous on the time axis, but in either case, the mutual data is not continuous on the time axis. May be. For example, the forecasting device 1 is based on a past frame (past data) consisting of 100- (t-2), 100- (t-3), ..., 100- (t-4). The prediction target data) may be virtually predicted. For example, the forecasting device 1 has 100- (t + 1), 100- (t-2), ..., 100- (t + 10) based on a past frame (past data) composed of 100- (t-3). A frame (future data) may be predicted. For example, the forecasting device 1 has a frame of 100− (t + 1) (based on a past frame (past data) composed of 100− (t−10), 100− (t−20) and 100− (t−30) ( Future data) may be predicted. The forecasting device 1 may forecast future data based on time-series data composed of past data with different time intervals. The data number determination device 2 of the first embodiment includes a virtual prediction signal generation unit 13 as a prediction unit, a virtual prediction error calculation unit 14 as a difference determination unit, and a virtual prediction error minimization determination unit as a data number determination unit 15. The virtual forecast signal generation unit 13 as the prediction unit predicts the prediction target data included in the time-series data based on a different number of past data among a plurality of past data included in the time-series data. Generate data for each piece. The virtual prediction error calculation unit 14 serving as a difference determination unit determines the difference between the prediction target data having temporal correlation with the past data and the prediction data for each number. The virtual forecast error minimization determination unit 15 as the data number determination unit determines the number of past data (p ^* ) based on the difference determined for each number. Thereby, the data number determination device 2 of the first embodiment can determine the number of data to be referred to when data is compressed. In addition, the data number determination device 2 of the first embodiment can predict future data at a specific spatiotemporal position based on the determined number of data.

The forecasting device 1 of the first embodiment further includes a forecast signal generation unit 19 as a signal generation unit. The forecast device 1 of the first embodiment may include a forecast signal generation unit 19 as a part of the data number determination device 2. Warning signal generator 19, the number ^{(p *)} on the basis of only one small number ^(p * -1) of the frame 100 as historical data and one frame 100-t than the time-series data A frame 100- (t + 1) is generated as future data having time correlation. Thereby, the data number determination device 2 of the first embodiment can determine the number of frames to be referred to when the frames are compressed.

  The forecasting device 1 according to the first embodiment predicts data that does not exist in time-series data based on a predetermined number of data that has a correlation with data that exists in time-series data and data that exists in time-series data. . In this case, the data number determination device 2 of the first embodiment can appropriately determine the predetermined number of data used by the forecasting device 1 for forecasting processing.

The virtual forecast signal generation unit 13 as the prediction unit generates prediction data of the prediction target data for each number of past data based on the prediction coefficient of the past data immediately before the prediction target data. The virtual forecast error minimization determination unit 15 as the data number determination unit determines the number of past data that minimizes the difference. The forecast signal generation unit 19 as a signal generation unit is based on the number of past data and the prediction target data that are one less than the number of past data (p ^* ) that minimizes the difference (p ^* -1). Generate future data. Thereby, the data number determination device 2 of the first embodiment can efficiently determine the number of frames to be referred to when compressing frames.

  The forecasting device 1 according to the first embodiment sets an appropriate forecast order p in forecasting processing based on a linear model related to blocks in a frame for time-series spatial information data, so that forecast accuracy can be improved. Become. The data compression apparatus according to the first embodiment can efficiently compress spatial information data using a prediction model that uses the correlation in the time direction of spatial information data that is time-series data.

(Second Embodiment)
The second embodiment is different from the first embodiment in that the optimum forecast order is determined by extending to the bit units constituting the frame. Only the differences from the first embodiment will be described below.
FIG. 6 is a diagram illustrating an example of the configuration of the forecasting device 1a. The forecast device 1a includes an input signal storage unit 10, an order update unit 11, a prediction coefficient calculation unit 12a, a virtual forecast signal generation unit 13a, a virtual forecast error calculation unit 14a, and a virtual forecast error minimization determination unit 15a. , Virtual forecast error minimum value storage unit 16, prediction coefficient determination unit 18a, forecast signal generation unit 19a, bit separation processing unit 20, virtual forecast error minimization determination unit 21, boundary bit update unit 22, virtual A prediction error minimum value storage unit 23, an optimum boundary bit storage unit 24, an optimum order / optimum boundary bit storage unit 25, and a bit separation processing unit 26 are provided.

  The forecast device 1a includes an order update unit 11, a prediction coefficient calculation unit 12a, a virtual forecast signal generation unit 13a, a virtual forecast error calculation unit 14a, a virtual forecast error minimization determination unit 15a, and a virtual forecast error minimum value storage. Unit 16, prediction coefficient determination unit 18a, prediction signal generation unit 19a, bit separation processing unit 20, virtual prediction error minimization determination unit 21, boundary bit update unit 22, and virtual prediction error minimum value storage unit 23 And the optimum boundary bit storage unit 24 is provided as a data number determination device 2a which is a device for determining the number of data. The prediction device 1a includes at least one of the input signal storage unit 10, the prediction coefficient determination unit 18a, the prediction signal generation unit 19a, the optimal order / optimum boundary bit storage unit 25, and the bit separation processing unit 26. You may provide as a function part of the data number determination apparatus 2a.

  Part or all of each functional unit of the forecasting device 1a is realized by, for example, a processor such as a CPU executing a program stored in the storage unit. The storage unit includes, for example, a nonvolatile recording medium (non-temporary recording medium) such as a magnetic hard disk device or a semiconductor storage device. A part or all of each functional unit may be realized using hardware such as an LSI or an ASIC, for example.

  The bit separation processing unit 20 reads a plurality of frames 100 stored in the input signal storage unit 10 as time series data, and separates the read time series data at positions determined by boundary bits by the boundary bit update unit 22. Thus, the upper bit signal and the lower bit signal are separated. The boundary bit represents a bit position serving as a reference for separating the upper bit signal and the lower bit signal. For example, the bit separation processing unit 20 separates a portion from the beginning of the frame 100 to a position determined by the boundary bit as an upper bit signal and a portion other than the upper bit signal as a lower bit signal.

  The prediction coefficient calculation unit 12a includes an upper bit prediction coefficient calculation unit 121 and a lower bit prediction coefficient calculation unit 122. The upper bit prediction coefficient calculation unit 121 reads an upper bit signal separated from one or more frames 100 determined by the prediction order p and the boundary bit as an upper bit reference frame. The higher-order bit prediction coefficient calculation unit 121 calculates a first p-order prediction coefficient that minimizes a prediction error when predicting the immediately preceding frame 100- (t-1) using the read higher-order bit reference frame. . The lower bit prediction coefficient calculation unit 122 reads a lower bit signal separated from one or more frames 100 determined by the prediction order p and the boundary bit as a lower bit reference frame. The lower-order bit prediction coefficient calculation unit 122 calculates a second p-order prediction coefficient that minimizes a prediction error when the immediately preceding frame 100- (t-1) is predicted using the read lower-bit reference frame. .

  The virtual forecast signal generator 13 a includes an upper bit virtual forecast signal generator 131 and a lower bit virtual forecast signal generator 132. The upper bit virtual forecast signal generation unit 131 reads the upper bit signal separated from the one or more frames 100 determined by the forecast order p and the boundary bits and the first p-order prediction coefficient, and reads the frame 100-t A first virtual forecast signal representing the result of the virtual forecast process is generated. The lower bit virtual forecast signal generation unit 132 reads the lower bit signal separated from the one or more frames 100 determined by the forecast order p and the boundary bits, and the second p-order prediction coefficient, and outputs the frame 100-t. A second virtual forecast signal representing the result of the virtual forecast process is generated.

  The virtual forecast error calculation unit 14 a includes an upper bit virtual forecast error calculation unit 141 and a lower bit virtual forecast error calculation unit 142. The upper bit virtual prediction error calculation unit 141 calculates a first virtual prediction error that is a prediction error between the upper bit signal of the frame 100-t and the first virtual prediction signal. The lower bit virtual prediction error calculation unit 142 calculates a second virtual prediction error that is a prediction error between the lower bit signal of the frame 100-t and the second virtual prediction signal.

The virtual forecast error minimizing determination unit 21 includes a virtual forecast error sum obtained based on the first virtual forecast error and the second virtual forecast error, and a virtual forecast error stored in the virtual forecast error minimum value storage unit 23. Compare the magnitude of the forecast error with the provisional minimum.
The boundary bit update unit 22 sequentially updates the boundary bits as B1 = 1,. The virtual forecast error minimum value storage unit 23 stores a provisional minimum value of the virtual forecast error. The optimum boundary bit storage unit 24 stores optimum boundary bits. The virtual forecast error minimizing determination unit 15 a is configured to calculate a provisional minimum value stored in the virtual forecast error minimum value storage unit 23 and a provisional minimum value of the virtual forecast error stored in the virtual forecast error minimum value storage unit 16. Compare sizes.

The optimum order / optimum boundary bit storage unit 25 stores the optimum forecast order p and the optimum boundary bit.
The bit separation processing unit 26 sets the optimal prediction order p and the optimal boundary bit stored in the optimal order / optimum boundary bit storage unit 25 as the prediction order and boundary bit of the prediction process, and the frame 100-t. A reference frame determined by the prediction order is read, and the read frame is separated at a position determined by a boundary bit to separate the upper bit signal and the lower bit signal.

The prediction coefficient determination unit 18a includes an upper bit prediction coefficient determination unit 181 and a lower bit prediction coefficient determination unit 182. The upper bit prediction coefficient determination unit 181 reads, as an upper bit reference frame, an upper bit signal separated from one or more frames 100 determined by the prediction order p and boundary bits. Upper bit prediction coefficient determining unit 181 uses the upper bits reference frames read, calculates a first prediction coefficients for the frame 100-t so as to minimize the prediction error E _p. The lower bit prediction coefficient determination unit 182 reads a lower bit signal separated from one or more frames 100 determined by the prediction order p and the boundary bit as a lower bit reference frame. Lower bit prediction coefficient determining unit 182, by using a lower bit reference frames read, calculates a second prediction coefficients of the frame 100-t so as to minimize the prediction error E _p.

  The prediction signal generation unit 19a includes an upper bit prediction signal generation unit 191, a lower bit prediction signal generation unit 192, and a bit synthesis processing unit 193. The upper bit prediction signal generation unit 191 uses the upper bit signal separated from the one or more frames 100 determined by the prediction order p and the boundary bit and the first prediction coefficient to generate a frame 100- A first forecast signal of (t + 1) is generated. The lower bit prediction signal generation unit 192 generates a frame 100- that is a prediction target frame based on the lower bit signal separated from the one or more frames 100 determined by the prediction order p and the boundary bit and the second prediction coefficient. A second forecast signal of (t + 1) is generated. The bit composition processing unit 193 generates a forecast signal by integrating the first forecast signal and the second forecast signal.

Hereinafter, the process of the forecast device 1a in the second embodiment will be described in detail.
[Adaptive forecast processing based on bit separation]
A prediction process based on separation in the bit depth direction will be described. Let f (x, t) be a signal having a bit depth of B [bits / sample]. The bit separation processing unit 20 converts the signal of f (x, t) into the upper bit signal f ^(B ₁ ^{+1: B)} (x, t) and the lower bit signal f ^{(1: B} ₁ ⁾ (x, t). To separate. Significant bit signal _{^{f (B 1 +1: B)}} (x, t) _is the upper _{B-B 1} bits from _B 1 +1 bit to the most significant bit. The lower bit signal f ^{(1: B} ₁ ⁾ (x, t) is the lower B ₁ bit from the least significant bit to the B ₁ bit.

  The bit composition processing unit 193 generates a forecast signal representing the result of the forecast process of the frame 100- (t + 1), which is the forecast target frame in the forecast process, as shown in Expression (20).

A _p ^(B ₁ ^{+1: B)} (t) and a _p ^{(1: B} ₁ ⁾ (t) in Expression 20 are coefficients obtained as solutions of Expressions (4) to (6). That is, a _p ^(B ₁ ^{+1: B)} (t) and a _p ^{(1: B} ₁ ⁾ (t) are calculated through minimizing the prediction error of the frame 100-t.
Note that the forecast signal to f ^(B ₁ ^{+1: B)} (x, t + 1, a _p ^(B ₁ ^{+1: B)} (t)) for the upper bit signal is in the range [2 ^B ₁ ⁺¹ , 2 ^B −1]. Is a value that should exist. “˜f” means to be written on “˜” f.

If the forecast signal for the upper bit signal exceeds the equivalence range, the forecast signal is clipped so that it falls within the equivalence range. That ^forecast signals below ₂ ^{B 1 +1} is set to ² _B ^{1 +1,} forecast signal exceeding the _{2 B} 1 -1 is a ² _B 1 -1. The forecast signal ~ f ^{(1: B} ₁ ⁾ (x, t + 1, _ap ^{(1: B} ₁ ⁾ (t)) for the lower-order bit signal is a value that should exist in the range [0,2 ^B ₁ -1]. is there. When the forecast signal for the lower bit signal exceeds the equivalence range, the forecast signal is clipped so that it falls within the equivalence range. Forecast values less than 0 are set to ^0, forecast signal exceeding the _{2 B} 1 -1 is a ² _B 1 -1.

[Adaptive bit separation]
A method for setting the boundary bit that becomes the boundary between the upper bit signal and the lower bit signal will be described. Prediction coefficient a _p ^(B ₁ ^{+1: B)} (t) for the upper bit signal set to minimize the prediction error in frame 100-t and prediction coefficient a _p ^{(1: B} ₁ ⁾ (t for the lower bit ) Is used to perform prediction processing for frame 100- (t + 1). At this time, the prediction error with respect to the frame 100- (t + 1) is expressed as Expression (21).

The bit separation processing unit 26 sets the optimum value of the boundary bit B ₁ based on the equation (22).

Prediction coefficient calculation unit 12a, the candidate value 1 of _{B 1,} ..., the prediction coefficients for B, and minimizes the prediction error of the frame ^{_{100-t a p (1:}} B 1) (t) and _a ^{p (B} ₁ ^{+ 1: B)} (t) is calculated. Further, the virtual prediction error calculation unit 14a uses the prediction coefficients a _p ^{(1: B} ₁ ⁾ (t) and a _p ^(B ₁ ^{+1: B)} (t) to generate a frame 100 that is a prediction target frame in the prediction process. Calculate a forecast error when performing forecast processing for-(t + 1). Finally, selecting a candidate value of B ₁ that minimizes the forecast error. Note that B ₁ = B means that bit separation is not performed. When the bit separation is not performed, the forecast signal generation unit 19a performs the forecast process based on the [forecast process].

  When simultaneously optimizing the boundary bit and the prediction order, the bit separation processing unit 26 sets the boundary bit and the prediction order based on Expression (23).

Here, B ₁ ^* and p ^* are selected from B × P candidates. P is the maximum value of the forecast order and is given from the outside. The above is a case where bit separation is performed into two ranges of an upper bit signal and a lower bit signal. Generalize this and consider the case of separation into m ranges. The possible values of m are m = 0,. The j-th boundary bit in the case of separation into the m range is represented as B _{m, j} . Here, j = 1,..., M. In addition, boundary bits in the case of separation into m ranges are collectively expressed as B _m = {B _{j, m} | j = 0,..., M}. Note that B _{m, 0} = 0 in order to obtain a unified notation.

In the case of separation into m ranges, the prediction coefficient a _p ^(B _{j−1, m} ^{: B} _{j, M} ⁾ (t) for the j-th range set to minimize the prediction error for the frame 100-t is expressed as It is assumed that the prediction process for the frame 100- (t + 1) is performed.
In the case of separation into m ranges, the coefficients for each range set to minimize the prediction error for the frame 100-t are collectively expressed as Expression (24).

  At this time, the prediction error for the frame 100- (t + 1) is expressed as Expression (25).

When separating into m ranges, the optimum value of the boundary bit B _m is set based on the equation (26).

  The optimum value of the number of ranges to be separated is set based on Expression (27).

Therefore, the optimum number of ranges is m ^* , and the optimum boundary bit at that time is B ^* _m ^* .

  When the boundary bit and the prediction order are simultaneously optimized, the boundary bit and the prediction order are set based on the equation (28) for each number of ranges.

  The optimum value of the number of ranges to be separated is set based on Expression (29).

Therefore, the optimum range number is m ^* , the optimum boundary bit at that time is B ^* _m ^* , and the optimum forecast order at that time is p ^* _m ^* .

Next, an example of the operation of the forecasting device 1a will be described.
FIG. 7 is a flowchart showing an example of the operation of the forecasting device 1a.
The input signal storage unit 10 reads a plurality of frames 100 as time series data (step S201). The order update unit 11 updates the predicted order p in the order of predicted order candidates p = 1,..., P (step S202). By executing step S203 to step S209, the data number determination device 2a determines the optimal forecast order p. The boundary bit update unit 22 updates the boundary bit B1 in the order of boundary bit candidates B1 = 1,..., B (step S203). For example, the boundary bit update unit 22 updates the boundary bits in order. The data number determination device 2a determines the optimum boundary bit by executing the optimum boundary bit update process (step S204). A specific process of the optimum boundary bit update process will be described later.

  When the optimum boundary bit update processing is executed for all the boundary bits (step S205), the virtual prediction error minimization determination unit 15a and the virtual prediction error minimum value storage unit obtained based on the boundary bits 16 is compared with the provisional minimum value stored in 16 (step S206). When the virtual forecast error sum is equal to or greater than the provisional minimum value (step S206: NO), the virtual forecast error minimization determination unit 15a advances the process to step S209. For example, the virtual forecast error minimization determination unit 15a instructs the order update unit 11 to update the forecast order.

When the virtual forecast error sum is smaller than the provisional minimum value (step S206: YES), the virtual forecast error minimization determination unit 15a sets the smaller virtual forecast error as the provisional minimum value of the virtual forecast error, and stores the virtual forecast error minimum value. The value stored in the unit 16 is updated (step S207). Further, the virtual forecast error minimization determination unit 15a updates the forecast order stored in the optimum order / optimum boundary bit storage unit 25 with the current forecast order p, and the optimum boundary bit for the updated forecast order p ^*. Are recorded in the optimum order / optimum boundary bit storage unit 25 (step S208).

When the processing of step S203 to step S208 is executed for all forecast orders (step S209), the bit separation processing unit 26 predicts the forecast order p ^* and the boundary bits stored in the optimum order / optimum boundary bit storage unit 25. Is set to the prediction order and boundary bits used in the prediction processing of the frame 100- (t + 1), which is the prediction target frame (step S210).

The bit separation processing unit 26 separates the reference frame determined by the set prediction order into an upper bit signal and a lower bit signal using the set boundary bits (step S211). The upper bit prediction coefficient determination unit 181 calculates the first prediction coefficient of the frame 100-t based on the upper bit signal of the reference frame determined by the frame 100-t and the prediction order p ^* (Step S212). The upper bit prediction signal generation unit 191 generates the first prediction signal of the frame 100- (t + 1), which is the prediction target frame, based on the upper bit signal of the reference frame determined by the frame 100-t and the prediction order p ^*. (Step S213).

The lower bit prediction coefficient determination unit 182 calculates the second prediction coefficient of the frame 100-t based on the lower bit signal of the reference frame determined by the frame 100-t and the prediction order p ^* (Step S214). The lower bit prediction signal generation unit 192 generates a second prediction signal of the frame 100- (t + 1), which is the prediction target frame, based on the lower bit signal of the reference frame determined by the frame 100-t and the prediction order p ^*. (Step S215). Thereafter, the bit synthesis processing unit 193 synthesizes the first forecast signal and the second forecast signal to generate a forecast signal of the frame 100- (t + 1) that is the forecast target frame (step S216).

FIG. 8 is a flowchart showing an example of the operation of the forecasting device 1a. In addition, in FIG. 8, the optimal boundary bit update process which the forecast apparatus 1a performs is demonstrated.
The bit separation processing unit 20 reads a plurality of frames 100 stored in the input signal storage unit 10 as time series data, and separates the read time series data into upper bit signals and lower bit signals by boundary bits. (Step S2041). Based on the upper bit signal of the reference frame determined by the frame 100-t and the prediction order p ^* , the upper bit prediction coefficient calculation unit 121 minimizes the prediction error when the frame 100- (t−1) is the prediction target frame. The first p-th order prediction coefficient to be converted is calculated (step S2042). The upper bit virtual prediction signal generation unit 131 reads the upper bit signal of the reference frame determined by the frame 100-t and the prediction order p ^* and the first p-order prediction coefficient, and performs virtual prediction processing of the frame 100-t. A first virtual forecast signal representing the result is generated (step S2043).

The upper bit virtual prediction error calculation unit 141 reads the frame 100-t and the generated first virtual prediction signal, and calculates the first virtual prediction error of the frame 100-t (step S2044). The lower bit prediction coefficient calculation unit 122 minimizes the prediction error when the frame 100- (t−1) is the prediction target frame based on the lower bit signal of the reference frame determined by the frame 100-t and the prediction order p ^*. The second p-th order prediction coefficient to be converted is calculated (step S2045). The lower bit virtual prediction signal generation unit 132 reads the lower bit signal of the reference frame determined by the frame 100-t and the prediction order p ^* and the second p-order prediction coefficient, and performs virtual prediction processing of the frame 100-t. A second virtual forecast signal representing the result is generated (step S2046). The lower bit virtual prediction error calculation unit 142 reads the frame 100-t and the generated second virtual prediction signal, and calculates the second virtual prediction error of the frame 100-t (step S2047).

The virtual forecast error minimizing determination unit 21 reads the first virtual forecast error and the second virtual forecast error, and calculates a virtual forecast error sum (step S2048). Thereafter, the virtual forecast error minimization determination unit 21 compares the calculated virtual forecast error sum with the provisional minimum value of the virtual forecast error stored in the virtual forecast error minimum value storage unit 23 (step S2049). When the virtual forecast error sum is smaller than the provisional minimum value (step S2049: YES), the virtual forecast error minimizing determination unit 21 is the calculated virtual forecast error sum and stored in the virtual forecast error minimum value storage unit 23. The provisional minimum value of the virtual forecast error is updated (step S2050). Further, the virtual prediction error minimization determination unit 21 updates the optimal boundary bit stored in the optimal boundary bit storage unit 24 to the current boundary bit (step S2051).
On the other hand, when the virtual forecast error sum is larger than the provisional minimum value of the virtual forecast error (step S2049: NO), the virtual forecast error minimizing determination unit 21 instructs the boundary bit update unit 22 to update the boundary bit. .

  As described above, the data number determination device according to the second embodiment includes the bit separation processing unit 20 as the separation processing unit, the virtual forecast signal generation unit 13a as the prediction unit, and the virtual forecast error calculation unit as the difference calculation unit. 14a and a virtual forecast error minimizing determination unit 15a as a data number determination unit. The bit separation processing unit 20 separates each of a plurality of past data, which is included in the time series data and is referred to when data is compressed, into a plurality of data. The virtual forecast signal generation unit 13a generates prediction data for each combination of the number of past data used for processing and the separated data. The virtual forecast error calculation unit 14a calculates a process for calculating the difference between the prediction target data having a temporal correlation with the past data and the predicted data for each of the separated data, for each candidate of the number of past data used for the process. To do. The virtual forecast error minimizing determination unit 15a determines the number of past data based on the difference obtained for each number of candidates. Thereby, in the prediction process for the block-based linear model, it is possible to set the optimum boundary bit and the prediction order, and to improve the prediction accuracy.

  In addition, as a breakdown of the 8-bit processing unit, there are many bits based on data that are completely unrelated to the upper bit signal 3 bits and the lower bit signal 5 bits, and only the processing unit is considered when performing prediction or forecasting. Doing so may increase the prediction residual and the amount of calculation. The data number determination device according to the second embodiment generates a forecast signal by separating bits at an optimum bit boundary in consideration of the above situation. As a result, it is possible to prevent an increase in the prediction residual and the calculation amount.

  As described above, in the data number determination device according to the second embodiment, the virtual prediction error minimization determination unit 15a as the data number determination unit determines the number of past data that minimizes the sum of the differences between the separated data. decide. Thereby, the data number determination apparatus of the second embodiment can efficiently determine the number of frames to be referred to when compressing frames.

  As described above, the data number determination device according to the second embodiment further includes the virtual forecast error minimization determination unit 21 as a separation position determination unit that determines a position where past data is separated. The virtual forecast error minimization determination unit 21 determines a separation position where the sum of the differences between the separated signals obtained when the past data is separated with the position candidates to be separated is minimized. Thereby, the data number determination apparatus of the second embodiment can efficiently determine the optimum separation position for determining the number of frames to be referred to when compressing frames.

(Third embodiment)
The third embodiment is different from the first embodiment and the second embodiment in that the optimal forecast order is determined by extending to a class composed of a plurality of frames. Further, the third embodiment is different from the first embodiment and the second embodiment in that (third exception processing) in [exception processing] is not executed. Hereinafter, only differences from the first embodiment and the second embodiment will be described.

  FIG. 9 is a diagram illustrating an example of the configuration of the forecasting device 1b. The forecast device 1b includes an input signal storage unit 10b, an order update unit 11, a prediction coefficient calculation unit 12b, a virtual forecast signal generation unit 13b, a virtual forecast error calculation unit 14b, and a virtual forecast error minimization determination unit 15b. The virtual forecast error minimum value storage unit 16, the prediction coefficient determination unit 18b, the forecast signal generation unit 19b, the class separation processing unit 28, the virtual forecast error minimization determination unit 29, the reference pattern update unit 30, and the virtual A prediction error minimum value storage unit 31, an optimum reference pattern storage unit 32, and an optimum order / optimum reference pattern storage unit 33 are provided.

  The forecast device 1b includes an input signal storage unit 10b, an order update unit 11, a prediction coefficient calculation unit 12b, a virtual forecast signal generation unit 13b, a virtual forecast error calculation unit 14b, and a virtual forecast error minimization determination unit 15b. , Virtual forecast error minimum value storage unit 16, class separation processing unit 28, virtual forecast error minimization determination unit 29, reference pattern update unit 30, virtual forecast error minimum value storage unit 31, and optimum reference pattern storage unit 32 as a data number determination device 2b which is a device for determining the number of data. The prediction device 1b may include at least one of the prediction coefficient determination unit 18b, the prediction signal generation unit 19b, and the optimal order / optimum reference pattern storage unit 33 as a functional unit of the data number determination device 2b. Moreover, the prediction coefficient calculation unit 12b, the virtual forecast signal generation unit 13b, the virtual forecast error calculation unit 14b, the virtual forecast error minimization determination unit 29, and the reference pattern update unit 30 determine the number of data. You may comprise as a class determination apparatus which is an apparatus which determines a class.

  Part or all of each functional unit of the forecasting device 1b is realized by, for example, a processor such as a CPU executing a program stored in the storage unit. The storage unit includes, for example, a nonvolatile recording medium (non-temporary recording medium) such as a magnetic hard disk device or a semiconductor storage device. A part or all of each functional unit may be realized using hardware such as an LSI or an ASIC, for example.

  The class separation processing unit 28 separates the input frame into a plurality of (for example, C) classes according to the attribute. Examples of attributes include gender, age group, etc. in the case of human flow data. In the following description, an attribute is called a class. That is, the class separation processing unit 28 separates the input frame into classes corresponding to the attributes of the input frame.

The input signal storage unit 10b stores a plurality of frames as time series data. For example, the input signal storage unit 10b stores a plurality of frames for each class in association with the class ID. The class ID is an ID for identifying a class.
The reference pattern update unit 30 sequentially updates the reference pattern. The reference pattern represents a class ID and a reference frame. That is, the reference pattern update unit 30 sequentially updates the class and the reference frame corresponding to the class.

  The prediction coefficient calculation unit 12b reads from the input signal storage unit 10b the frame 100- (t-1) immediately before the frame 100-t corresponding to the class ID updated by the reference pattern update unit 30. The prediction coefficient calculation unit 12b reads one or more frames 100 determined by the prediction order p updated by the order update unit 11 from the input signal storage unit 10b as reference frames. The prediction coefficient calculation unit 12b calculates a p-th order prediction coefficient that minimizes a prediction error when the immediately preceding frame 100- (t-1) is predicted using the reference frame.

The virtual forecast signal generation unit 13b reads one or more reference frames determined by the forecast order p and the reference pattern, and the p-th order prediction coefficient calculated by the prediction coefficient calculation unit 12b, and performs virtual prediction processing of the frame 100-t. A virtual forecast signal representing the result of is generated.
The virtual forecast error calculating unit 14b calculates a virtual forecast error that is a forecast error between the frame 100-t obtained from the input signal storage unit 10b and the virtual forecast signal representing the frame 100-t.

  The virtual forecast error minimizing determination unit 29 determines the magnitudes of the virtual forecast error calculated by the virtual forecast error calculation unit 14 b and the provisional minimum value of the virtual forecast error stored in the virtual forecast error minimum value storage unit 31. Compare. The virtual forecast error minimizing determination unit 29 determines the optimum reference pattern that minimizes the virtual forecast error by comparing the magnitude of the virtual forecast error for each reference pattern.

The virtual forecast error minimum value storage unit 31 stores a provisional minimum value of the virtual forecast error. The optimum reference pattern storage unit 32 stores an optimum reference pattern. The virtual forecast error minimizing determination unit 15b provisionally minimizes the virtual forecast error stored in the virtual forecast error minimum value storage unit 31 and the virtual forecast error provisional stored in the virtual forecast error minimum value storage unit 16. Compare the size with the minimum value. The optimum order / optimum reference pattern storage unit 33 stores the optimum forecast order p ^* and the optimum reference pattern.

The prediction coefficient determination unit 18b sets the forecast order p ^* and the reference pattern stored in the optimum order / optimum reference pattern storage unit 33 as the forecast order and reference pattern of the forecast process, and sets the frame 100-t and the forecast order p. ^* and on the basis of the reference frame determined by the reference pattern, and calculates prediction coefficients of the frame 100-t so as to minimize the prediction error E _p.
The prediction signal generation unit 19b generates a frame 100- (t + 1) that is a prediction target frame based on one or more reference frames determined by the prediction order p ^* and the reference pattern and the prediction coefficient calculated by the prediction coefficient determination unit 18b. ) Forecast signal.

  10 and 11 are diagrams illustrating an example of a forecast process based on a class. 10 and 11, the frame at the current time among the plurality of frames 100 is a frame 100-t. The time interval between the frames 100 is a unit time length such as 1 minute, for example. As shown in FIGS. 10 and 11, the forecasting device 1b according to the third embodiment classifies the frames 100 based on attributes, and performs forecasting processing for each class.

  10 and 11, three classes of Class 1 (for example, women in their 30s), Class 2 (for example, women in their 40s), and Class 3 (for example, women in their 50s) are exemplified as Class. In FIG. 10, the forecasting device 1 b performs class 2 forecast processing in the same class frame 100-t, frame 100-(t + 1-D), another class frame 100-t, and frame 100-(t + 1-). D) is used as a reference frame. FIG. 11 shows a case where the forecasting device 1b uses more frames 100 than FIG. 10 as reference frames in the Class2 forecasting process.

Hereinafter, the process of the forecast device 1b in the third embodiment will be described in detail.
[Inter-class prediction]
The class separation processing unit 28 decomposes the spatial information data f (x, t) for each attribute. When it is decomposed for each attribute, f (x, t) is expressed as the sum of the attributes as shown in Expression (30).

  The virtual forecast signal generation unit 13b generates a forecast signal of the frame 100-t belonging to the classes to c as shown in Expression (31). “˜c” means that “˜” is written on c.

Here, the i-th reference block is defined by the class ID c _i (t) and the time position l _i (t). The vector c _p (t) (= c ₁ (t),..., C _p (t)) is a class list of reference blocks. The vector l _p (t) (= l ₁ (t),..., L _p (t)) is a frame list of reference blocks.

The possible ranges of the scalars c _i (t), l _i (t) are c _i (t) = 0,..., C−1, l _i (t) = 0,..., L−1, A, A ±. 1,..., A ± L,..., MA, MA ± 1,. Here, C is the number of classes. L is a reference frame existing section in the vicinity. A is a time interval corresponding to a signal cycle, M is a parameter given from the outside, and means that a frame before the first cycle to the previous M cycle and its neighboring frames are used as reference candidates. Examples of cycles include one day, one week, one month, one year, etc. A specific list construction method will be described later.

A vector a _p (t) (= a ₀ (t), a ₁ (t),..., A _p (t)) is obtained as a solution of the equation (1). The prediction coefficient calculation unit 12b derives a solution by replacing f (x, t−i) in the same expression with g (x, t−l _i ). In other words, the prediction coefficient a _p (t) refers to p prediction reference blocks defined by a plurality of lists (vector c _p (t) and vector l _p (t)), and predicts the frame 100-t. It is set based on a standard that minimizes the error. This prediction coefficient a _p (t) is used as a prediction coefficient of the frame 100-t as shown in Expression (31).

Next, for the purpose of generalizing the forecasting process, when predicting the ˜τ frame (frame 100− (t + ˜τ)) using the coefficient a _p (τ) that minimizes the prediction error of the τ frame. think of. “˜τ” means that “˜” is described above “τ”. This is the case where the forecast coefficient a _p (τ) is generated based on the forecast process of the τ-th frame while the forecast process is applied to the τ-th frame. In this case, the virtual forecast error calculation unit 14b calculates a forecast error based on Expression (32).

[Adaptive selection of forecast order]
A method for setting the forecast order p, which is the order used in the forecast process of Equation (32) (method for determining the number of data) will be described. In the third embodiment, the method for estimating the optimum value of the predicted order is composed of two steps. In the first step, an optimal reference pattern is identified for the forecast process of order p. First, the prediction coefficient calculation unit 12b calculates, for each reference pattern, a p-th order prediction coefficient a _p (t−1) that minimizes a prediction error when the frame 100- (t−1) is a prediction target. calculate. Next, the virtual forecast error calculation unit 14 uses the prediction coefficient a _p (t−1) to obtain a p-th order forecast error when the frame 100-t is a forecast target. Further, the virtual forecast error minimizing determination unit 29 compares the forecast errors of the respective reference patterns and identifies the reference pattern that minimizes the forecast error. That is, the process of the first step is a process of selecting a reference pattern that minimizes the virtual prediction error with respect to the frame 100-t, and can be expressed as Expression (33).

In the second step, the optimal order is identified using the reference pattern obtained in the first step. By using the reference pattern obtained in the first step, the virtual prediction error in the order p can be minimized. Here, p = 1,. P is assumed to be given from the outside. Therefore, the virtual forecast error minimizing determination unit 15b compares the minimum virtual forecast errors in the respective orders, and identifies the order p that minimizes the virtual forecast error as the forecast order p ^* . That is, the process of the second step is a process of identifying the forecast order that minimizes the virtual forecast error with respect to the frame 100-t, and can be expressed as Expression (34).

The forecasting device 1b performs forecasting processing for each class based on the formula (31) using the forecast order p ^* defined by the above formula. In the optimization of the prediction order p ^* described above, the amount of calculation can be reduced by limiting the combinations of reference patterns. For example, there are the following methods.
1. For p = 1, identify the class ID and reference frame in the primary virtual forecast. In this case, equation (33) becomes equation (35).

2. for p = 2, ..., P
3. Under the condition of using p-1 class IDs and reference frames determined by the (p-1) th order virtual forecast, the lowest pth class ID and reference frame are determined based on Expression (36).

4). Vector ^{_{c p * (t-1)}} , l p * (t-1) the p-1 component from a first vector ^{_{c p * (t-1)}} , inherit * vector _l ^p (t-1) , C _p (t−1), and l _p (t−1) are set as scalar c _p ^* (t−1) and scalar l _p ^* (t−1).
If the said method is followed, it will become possible to limit the reference pattern with respect to each forecast order to one set.

Next, an example of the operation of the forecasting device 1b will be described.
FIG. 12 is a flowchart illustrating an example of the operation of the forecasting device 1b.
The class separation processing unit 28 reads the plurality of frames 100 as time series data (step S301). The class separation processing unit 28 separates the plurality of frames 100 that are input time-series data into a plurality of classes according to attributes (step S302). The class separation processing unit 28 records the separated frame 100 in 10b in association with the class ID.

  The order update unit 11 updates the forecast order p in the order of forecast order candidates p = 1,..., P (step S303). By executing step S304 to step S310, the data number determination device 2b determines the optimal forecast order p. The reference pattern update unit 30 updates the reference patterns in order (step S304). For example, the reference pattern update unit 30 updates the class IDs in order. The data number determination device 2b determines an optimal optimum pattern by executing an optimal reference pattern update process (step S305). The specific process of the optimum reference pattern update process will be described later.

  When the optimum reference pattern update process is executed for all reference patterns (step S306), the virtual forecast error minimizing determination unit 15b stores the virtual forecast error obtained in the reference pattern and the virtual forecast error minimum value storage unit 16. The size of the provisional minimum value is compared (step S307). When the virtual forecast error is greater than or equal to the provisional minimum value (step S307: NO), the virtual forecast error minimization determination unit 15b advances the process to step S310. For example, the virtual forecast error minimizing determination unit 15b instructs the order update unit 11 to update the forecast order.

On the other hand, when the virtual forecast error is smaller than the provisional minimum value of the virtual forecast error (step S307: YES), the virtual forecast error minimization determination unit 15b sets the smaller virtual forecast error as the provisional minimum value of the virtual forecast error, The value stored in the forecast error minimum value storage unit 16 is updated (step S308). Further, the virtual forecast error minimizing determination unit 15b updates the forecast order stored in the optimum order / optimum reference pattern storage unit 33 with the current forecast order p, and the optimum reference pattern for the updated forecast order p ^*. Are recorded in the optimum order / optimum reference pattern storage unit 33 (step S309).

When the processing of step S304 to step S309 is executed for all prediction orders (step S310), the prediction coefficient determination unit 18b stores the prediction order p ^* and the reference pattern stored in the optimal order / optimum reference pattern storage unit 33. Is set to the prediction order and reference pattern used in the prediction processing of the frame 100- (t + 1), which is the prediction target frame (step S311). The prediction coefficient determination unit 18b calculates a prediction coefficient of the frame 100-t based on the frame 100-t and the reference frame determined by the prediction order p ^* and the reference pattern (step S312). The prediction signal generation unit 19b generates a prediction signal for the frame 100- (t + 1), which is the prediction target frame, based on one or more reference frames determined by the prediction order p ^* and the reference pattern and the prediction coefficient (step S313). ).

FIG. 13 is a flowchart showing an example of the operation of the forecasting device 1b. In addition, in FIG. 13, the optimal reference pattern update process which the forecast apparatus 1b performs is demonstrated.
The prediction coefficient calculation unit 12b reads, as a reference frame, one or more frames 100 determined by the forecast order p and the reference pattern updated by the order update unit 11. The prediction coefficient calculation unit 12b calculates a p-th order prediction coefficient that minimizes a prediction error when the frame 100- (t-1) is a prediction target frame (step S3051). The virtual forecast signal generation unit 13b reads one or more reference frames determined by the forecast order p and the reference pattern, and the p-th order prediction coefficient calculated by the prediction coefficient calculation unit 12b, and performs virtual prediction processing of the frame 100-t. A virtual forecast signal representing the result is generated (step S3052). The virtual forecast error calculation unit 14b reads the frame 100-t and the generated virtual forecast signal, and calculates the virtual forecast error of the frame 100-t (step S3053).

  The virtual forecast error minimizing determination unit 29 determines the magnitudes of the virtual forecast error calculated by the virtual forecast error calculation unit 14 b and the provisional minimum value of the virtual forecast error stored in the virtual forecast error minimum value storage unit 31. Compare (step S3054). When the virtual forecast error is larger than the provisional minimum value (step S3054: NO), the virtual forecast error minimization determination unit 29 instructs the reference pattern update unit 30 to update the reference pattern.

  On the other hand, when the virtual forecast error is smaller than the provisional minimum value (step S3054: YES), the virtual forecast error minimization determination unit 29 sets the smaller virtual forecast error as the provisional minimum value of the virtual forecast error and sets the virtual forecast error minimum value. The value stored in the storage unit 31 is updated (step S3055). Further, the virtual forecast error minimizing determination unit 29 updates the reference pattern stored in the optimal reference pattern storage unit 32 with the current reference pattern as the optimal reference pattern (step S3056).

  As described above, the class determination device according to the third embodiment includes a virtual forecast signal generation unit 13b as a prediction unit, a virtual forecast error calculation unit 14b as a difference calculation unit, and a virtual prediction error minimization as a class determination unit. And a determination unit 29. The virtual forecast signal generation unit 13b includes a class different from the class to which the prediction target data referred to predict the prediction target data, which is included in the time-series data classified into the class based on the attribute, is predicted. Prediction data is generated for each class using the past data of at least one class classified as (1). The virtual forecast error calculation unit 14b performs a process of calculating the difference between the prediction target data having a temporal correlation with the past data and the prediction data for each class that has generated the prediction data. The virtual forecast error minimizing determination unit 29 determines a class with the smallest difference obtained for each class as a class for determining the number of a plurality of past data referred to predict the prediction target data. . Thereby, an appropriate reference class can be determined. Therefore, it is possible to set an appropriate reference class, a reference frame in each reference class, and a prediction order in a prediction process for a block-based linear model that refers to another class, and to improve prediction accuracy. .

  You may make it implement | achieve at least one part of the data number determination apparatus, forecast apparatus, and class determination apparatus in embodiment mentioned above with a computer. In that case, a program for realizing this function may be recorded on a computer-readable recording medium, and the program recorded on this recording medium may be read into a computer system and executed. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, the “computer-readable recording medium” dynamically holds a program for a short time like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory inside a computer system serving as a server or a client in that case may be included and a program held for a certain period of time. Further, the program may be a program for realizing a part of the above-described functions, and may be a program capable of realizing the functions described above in combination with a program already recorded in a computer system. You may implement | achieve using programmable logic devices, such as FPGA (Field Programmable Gate Array).

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes designs and the like that do not depart from the gist of the present invention.

  The present invention is applicable to a data number determination device that determines the number of data and a prediction device that predicts future data.

  DESCRIPTION OF SYMBOLS 1, 1a, 1b ... Forecast apparatus 2, 2a, 2b ... Data number determination apparatus, 10 ... Input signal memory | storage part, 11 ... Order update part, 12, 12a, 12b ... Prediction coefficient calculation part, 121 ... Upper bit prediction coefficient Calculation unit, 122 ... lower bit prediction coefficient calculation unit, 13, 13a, 13b ... virtual forecast signal generation unit, 131 ... upper bit virtual forecast signal generation unit, 132 ... lower bit virtual forecast signal generation unit, 14, 14a, 14b ... Virtual forecast error calculation unit, 141 ... upper bit virtual forecast error calculation unit, 142 ... lower bit virtual forecast error calculation unit, 15, 15a, 15b ... virtual forecast error minimization determination unit, 16 ... virtual forecast error minimum value storage unit, 17 ... Optimum order storage unit 18, 18a, 18b ... Prediction coefficient determination unit, 181 ... Upper bit prediction coefficient determination unit, 182 ... Lower bit prediction coefficient determination unit 19, 19a, 1 b ... prediction signal generation unit, 191 ... upper bit prediction signal generation unit, 192 ... lower bit prediction signal generation unit, 20 ... bit separation processing unit, 21 ... virtual prediction error minimization determination unit, 22 ... boundary bit update unit, 23 ... virtual forecast error minimum value storage unit, 24 ... optimal boundary bit storage unit, 25 ... optimal order / optimum boundary bit storage unit, 26 ... bit separation processing unit, 28 ... class separation processing unit, 29 ... virtual prediction error minimization determination , 30 ... Reference pattern update unit, 31 ... Virtual forecast error minimum value storage unit, 32 ... Optimal reference pattern storage unit, 33 ... Optimal order / optimum reference pattern storage unit, 100 ... Frame, 200 ... Frame group, 201 ... Frame Group, 202 ... frame group

Claims (5)

A separation processing unit that separates each of a plurality of past data into a plurality of data, which is referred to in order to predict the prediction target data included in the time series data,
A prediction unit that generates prediction data for each combination of the number of past data used for processing and the separated data;
The difference calculation is performed for each candidate of the number of past data used in the process, the process of calculating the difference between the prediction target data having a temporal correlation with the past data and the prediction data for each of the separated data. And
A data number determination unit for determining the number of the past data based on the difference obtained for each of the number candidates;
A data number determination device comprising:   The data number determination device according to claim 1, wherein the data number determination unit determines the number of the past data that minimizes a sum of differences between the separated data. A separation position determination unit for determining a position for separating the past data;
3. The separation position determination unit according to claim 1, wherein the separation position determination unit determines a separation position that minimizes a sum of differences between the signals after separation obtained when the past data is separated from candidates of positions to be separated. Data number determination device. A data number determination method executed by a data number determination device for determining the number of data,
A separation processing step for separating each of a plurality of past data into a plurality of data, which is referred to in order to predict the prediction target data included in the time series data,
A prediction step for generating prediction data for each combination of the number of past data used for processing and the separated data;
The difference calculation is performed for each candidate of the number of past data used in the process, the process of calculating the difference between the prediction target data having a temporal correlation with the past data and the prediction data for each of the separated data. Steps,
A data number determination step for determining the number of the past data based on the difference obtained for each of the number candidates;
A method for determining the number of data. On the computer,
A separation processing step for separating each of a plurality of past data into a plurality of data, which is referred to in order to predict the prediction target data included in the time series data,
A prediction step for generating prediction data for each combination of the number of past data used for processing and the separated data;
The difference calculation is performed for each candidate of the number of past data used in the process, the process of calculating the difference between the prediction target data having a temporal correlation with the past data and the prediction data for each of the separated data. Steps,
A data number determination step for determining the number of the past data based on the difference obtained for each of the number candidates;
Program for determining the number of data to execute.

Images