JP2006072691A - Image analysis device, image analysis program and image analysis method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an image analysis device, an image analysis method and an image analysis program for accurately extracting movement of an image while requiring a small amount of prior knowledge. <P>SOLUTION: An image acquisition part 11 acquires image data as a processing object, a pre-processing part 12 and a clustering part 13 use a Potts model to handle each pixel of the image data as a particle, and cluster the movement of the image in a super paramagnetic state of the particles by a statistical method, and a movement extraction part 14 extracts the cluster of the movement in the image based on the clustering result. Hereat, the clustering part 13, by explicitly using the movement of cluster, uses the interaction between defined pixels to cluster the movement of the image. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to an image analysis apparatus, an image analysis program, and an image analysis method for analyzing image motion.

  Extracting and classifying image motion in video is the core of image recognition technology and has a wide range of applications, and is essential for traffic monitoring, intruder detection, human recognition, and pre-processing of TV video segmentation. is there. It is also deeply related to a technique called Structure from Motion that estimates the three-dimensional shape of the target object from the motion of the image.

  Research on the analysis of such image movement has a long history, and has been known as optical flow analysis since the 1980s. For example, Non-Patent Document 1 describes a constraint called BCC (Brightness Constant Constraint) based on the brightness information of each pixel. It is disclosed that conditions are derived and motion estimation is performed.

  In addition, Non-Patent Document 2 describes, as a study of image analysis using the Potts model, a group of points interspersed in an image as magnetic particles and accumulation of local interactions between the points. Generating a secure cluster is disclosed. The Potts model is an extension of the Ising model to multiple values. In the Potts model, the result of the cluster differs by changing the temperature and local interaction. In this Non-Patent Document 2, the temperature is continuously changed, and the system changes from a ferromagnetic state to a superparamagnetic state and a paramagnetic state while detecting the system susceptibility, and the superparamagnetic state is detected. Concludes that the best cluster is obtained.

As described above, in order to use the Potts model for motion clustering in an image, it is necessary to define a local interaction. However, a motion cannot be uniquely determined only by BCC. The interaction definition used in the above cannot be used as is. Therefore, the present inventors, the luminance of the image I (x, y, t) , I (x, y, t) the partial differential of _{bf = (f x, f y} , f t), the pixel Assuming that the motion vector bu = (u, v) and the motion vector bu _i and bu _j of the pixel j adjacent to the pixel _i are the same, the corresponding constraint line is the same. Based on this, the similarity of motion between the pixel i and the pixel j is defined as d _ij = ‖bf _i × bf _j ‖ / ｆbf _i ‖‖bf _j 、, and the interaction between the pixels is calculated using this similarity Thus, a method for clustering image motion has been proposed (see Non-Patent Document 3).
B. Horn et al., "Determining optical flow", Artificial Intelligence, 1981, 17: 185-203. M. Blatt et al., “Clustering data through an analogy to the potts mode”, Neural Information Processing Systems, 1995, NIPS, Vol. 8, p. 416-p. 422 Keisuke Kinoshita et al., “Clustering of Moving Images Using Potts Model”, Information Processing Society of Japan Research Report, 2003, 2003-CVIM-138, p. 193-p. 200

  However, if the similarity of motion between the pixel i and the pixel j disclosed in Non-Patent Document 3 is used, the superparamagnetic state does not appear clearly, so that the motion of the image cannot be extracted with high accuracy. . In addition, as a conventional image motion analysis method, there is mainly a method in which a motion model is given in advance and a motion that conforms to the motion model is estimated. In this case, a motion model such as rigid body motion or motion smoothness constraint is used. It becomes necessary and prior knowledge becomes enormous.

  An object of the present invention is to provide an image analysis apparatus, an image analysis method, and an image analysis program capable of reducing prior knowledge and extracting an image motion with high accuracy.

  The image analysis apparatus according to the present invention includes an acquisition unit that acquires image data to be processed, and each pixel of the image data acquired by the acquisition unit using a spin model of a magnetic material in which a spin state is expanded to multiple values. Clustering means that uses a statistical method to cluster image motion in a state where particles are superparamagnetic, and an extraction means that extracts motion cluster in the image based on the clustering result of the clustering means. The means is to cluster the image motion using the interaction between pixels defined by explicitly using the cluster motion.

  In the image analysis apparatus according to the present invention, image data to be processed is acquired, and each pixel of the image data is treated as a particle using a spin model of a magnetic material in which the spin state is expanded to multiple values. The motion of the image in a magnetic state is clustered by a statistical method, and a cluster of motion in the image is extracted based on the clustering result. In this clustering process, the cluster motion is defined by using the cluster motion explicitly. Since the motion of the image is clustered using the interaction, the superparamagnetic state appears clearly, and the motion of the image in the superparamagnetic state can be clustered to extract the motion of the image with high accuracy. it can. Moreover, since global movements can be clustered simply by defining local interactions, prior knowledge can be reduced. As a result, prior knowledge can be reduced and the motion of the image can be extracted with high accuracy.

Clustering means, the brightness of the image I (x, y, t) , I (x, y, t) the partial differential bf = a _{(f x, f y, f} t), cluster pixel i belongs When the motion vector is BU _i = (U _i , V _i ), the particle i shown in the following equation (2) is used by using the motion similarity d _ij between the pixel i and the pixel j shown in the following equation (1). and (where, K is the average of the number of pixels in the neighborhood, a is the average of d _ij) interaction J _ij between the particle j that is calculated and clustering the movement of an image by using the calculated interaction J _ij preferable. In this case, the superparamagnetic state appears more clearly, and the image motion in the superparamagnetic state can be clustered to extract the image motion with higher accuracy.

  The clustering means preferably clusters the motion of the image using the Monte Carlo method. In this case, the clustering process can be performed at high speed.

The clustering means assigns a spin value to each pixel at random using the Swendsen-Wang algorithm, cuts or combines the link between the target pixel and the neighboring pixel according to the probabilities shown in the following formulas (3) and (4), and combines them. A group of pixels having the linked links is defined as a cluster, and a motion vector BU _i in each cluster is calculated, and a process of randomly assigning a spin value to each pixel so that the same spin value is assigned to pixels belonging to the same cluster is performed a predetermined number of times. It is preferable to cluster image motion by iterative processing. In this case, since the spin state is not randomly assigned to each pixel (particle) separately, but the same spin state can be assigned to the pixels belonging to the cluster, the clustering process can be performed at higher speed.

  The clustering means determines the temperature at which the magnetic susceptibility calculated using the inter-pixel interaction takes a value within a predetermined range when the temperature is swept as the temperature at which the particles are superparamagnetic. It is preferable to cluster image motion at temperature. In this case, since the temperature at which the superparamagnetic state is achieved can be determined accurately, the motion of the image can be extracted with higher accuracy.

  The image analysis program according to the present invention includes an acquisition unit that acquires image data to be processed, and each pixel of the image data acquired by the acquisition unit using a spin model of a magnetic material in which a spin state is expanded to multiple values. Clustering means to cluster image motion in a state where particles are superparamagnetic using a statistical method, and a computer to function as an extraction means to extract the motion cluster in the image based on the clustering result of the clustering means The clustering means clusters the motion of the image using the interaction between the pixels defined by explicitly using the motion of the cluster.

  An image analysis method according to the present invention is an image analysis method for analyzing image motion using a computer that functions as an acquisition unit, a clustering unit, and an extraction unit, and the acquisition unit acquires image data to be processed. A state in which the first step and the clustering unit treat each pixel of the image data acquired by the acquiring unit as a particle using a spin model of a magnetic material in which the spin state is expanded to multiple values, and the particle becomes superparamagnetic A second step of clustering the motion of the image in the statistical method by a statistical method, and a third step of the extraction means extracting a motion cluster in the image based on the clustering result by the clustering means. , Use the cluster motion explicitly to define the image motion using the inter-pixel interaction It is intended to include the step of rastering.

  According to the present invention, in the clustering process, since the motion of the image is clustered using the interaction between the pixels defined by explicitly using the motion of the cluster, the superparamagnetic state clearly appears and the superparamagnetic state appears. Since it is possible to extract image motion with high accuracy by clustering image motion in a magnetic state, global motion can be clustered simply by defining local interaction, so prior knowledge Can be reduced.

  Hereinafter, an image analysis apparatus according to an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a configuration of an image analysis apparatus according to an embodiment of the present invention.

  The image analysis apparatus shown in FIG. 1 includes a computer device such as a personal computer, and includes an input device 1, ROM (read only memory) 2, CPU (central processing unit) 3, RAM (random access memory) 4, I / F (interface) unit 5, external storage device 6, display device 7, and recording medium drive device 8. Each block is connected to an internal bus, and various data and the like are input / output between the blocks via this bus, and various processes are executed under the control of the CPU 3.

  The input device 1 includes a keyboard, a mouse, and the like, and is used for a user to input various operation commands. The ROM 2 stores a system program and the like in advance. The external storage device 6 is composed of a hard disk drive or the like, and stores an image analysis program and the like which will be described later. The CPU 3 reads an image analysis program or the like from the external storage device 6 and executes an image analysis process or the like described later to control the operation of each block. The RAM 4 is used as a work area for the CPU 3. The display device 7 includes a CRT (cathode ray tube), a liquid crystal display device, or the like, and displays various screens under the control of the CPU 3. The I / F unit 5 includes an image capture board and the like, and the image output apparatus 10 includes devices capable of outputting image data such as a television, a video camera, and a DVD recorder. The I / F unit 5 is connected to the image output device 10 and outputs image data output from the image output device 10 to the CPU 3 or the like.

  The image analysis program may be recorded on a computer-readable recording medium 9 composed of a CD-ROM, a DVD, or the like. In this case, the image analysis program read from the recording medium 9 is installed in the external storage device 6 using the recording medium driving device 8 constituted by a CD-ROM drive, a DVD drive or the like. In addition, when the image analysis apparatus has a communication device that controls communication with the outside, and an image analysis program or the like is stored in another computer or the like connected via the network, the computer or the like via the network Thus, an image analysis program or image data to be processed may be downloaded.

  Next, main functions of the image analysis apparatus configured as described above will be described. FIG. 2 is a main functional block diagram of the image analysis apparatus shown in FIG.

  As shown in FIG. 2, the image analysis apparatus functions as an image acquisition unit 11, a preprocessing unit 12, a clustering unit 13, a motion extraction unit 14, and an image processing unit 15 when the CPU 3 or the like executes an image analysis program or the like. To do.

  The image acquisition unit 11 includes the input device 1 and the I / F unit 5. For example, when the user inputs an image data acquisition command, the image acquisition unit 11 acquires image data to be processed from the image output device 10. The preprocessing unit 12 includes a CPU 3 and the like, and performs preprocessing such as smoothing processing and resolution conversion processing on image data to be processed, and converts the image data into image data having a resolution suitable for clustering processing.

The clustering unit 13 is configured by the CPU 3 and the like, and handles each pixel of the image data output from the preprocessing unit 12 as particles using a Potts model that is a spin model of a magnetic material in which the spin state is expanded to multiple values. The motion of images in a state where particles are superparamagnetic are clustered by a statistical method. At this time, the clustering unit 13 clusters the motion of the image using the interaction between pixels defined by explicitly using the motion of the cluster. For example, the clustering unit 13, the luminance of the image I (x, y, t) , I (x, y, t) the partial differential bf = a _{(f x, f y, f} t), the pixel i belongs When the motion vector of the existing cluster is BU _i = (U _i , V _i ), the following equation (2) is used by using the motion similarity d _ij between the pixel i and the pixel j shown in the following equation (1). The interaction J _ij (where K is the average of the number of neighboring pixels and a is the average of d _ij ) is calculated, and the motion of the image is clustered using the calculated interaction J _ij. To do. In this case, the superparamagnetic state appears more clearly, and the image motion in the superparamagnetic state can be clustered to extract the image motion with higher accuracy. Note that various values can be used as the number of neighboring pixels. For example, eight pixels positioned in the immediate vicinity of the upper, lower, left, and right sides of the target pixel can be used.

In the clustering process described above, the clustering unit 13 clusters image motion using the Monte Carlo method. Specifically, the clustering unit 13 uses the Swendsen-Wang algorithm (see R. Swendsen and JS. Wang. Nonuniversal critical dynamics in monte carlo simulation. Physics Review Letters, 58 (2): 86.88, 1987.). Then, a spin value is randomly assigned to each pixel, and the link between the target pixel and neighboring pixels is cut or combined according to the probability expressed by the following formulas (3) and (4), and a pixel group having the linked links is defined as a cluster. Then, the motion vector BU _i in each cluster is calculated, and the process of assigning the spin value randomly to each pixel is repeatedly processed a predetermined number of times so as to assign the same spin value to the pixels belonging to the same cluster, thereby clustering the motion of the image To do. In this case, since the spin state is not randomly assigned to each pixel, but the same spin state can be assigned to the pixels belonging to the cluster collectively, the clustering process can be performed at a higher speed.

  Note that the Monte Carlo method used for the clustering process is not particularly limited to the above example, and various modifications are possible. For example, a Markov chain Monte Carlo method such as Metropolis, Gibbs sampling may be used.

  In the clustering process described above, when the temperature is swept, the clustering unit 13 becomes superparamagnetic at a temperature at which the magnetic susceptibility calculated using the interaction between pixels takes a value within a predetermined range. The temperature of the state is determined, and image motion at the determined temperature is clustered. In this case, since the temperature at which the superparamagnetic state is achieved can be determined accurately, the motion of the image can be extracted with higher accuracy.

  The motion extraction unit 14 includes a CPU 3 and the like, and extracts a motion cluster in the image based on the clustering result of the clustering unit 13. The image processing unit 15 is composed of the CPU 3 and the like, and performs various image processing on the image data based on the extracted cluster. For example, traffic monitoring, intruder detection, person recognition, TV video clipping, etc. It performs image processing used for various applications. Note that the image processing unit 15 may be omitted when simply analyzing the motion of the image.

  In the present embodiment, the image acquisition unit 11 corresponds to an example of an acquisition unit, the preprocessing unit 12 and the clustering unit 13 correspond to an example of a clustering unit, and the motion extraction unit 14 corresponds to an example of an extraction unit.

Next, the Potts model used in this embodiment will be described in detail. FIG. 3 is a schematic diagram of the Ising model and the Potts model, and the arrows in the figure indicate the direction of spin. The Potts model shown in (b) of FIG. 3 changes the spin state of the Ising model, which is the spin model of the magnetic material shown in (a) of FIG. 3, from a binary (+1, −1) to a multivalue (1,. It is an extension to q) and is known as a spin model of a magnetic material (see FYWu. The potts model. Review of Modern Physics, 54 (1): 235-268, 1984., etc.). Each particle is assigned one of q states (1,..., Q). When the state of the spin assigned to the particle i is represented by σ _i , the Hamiltonian of the system is represented by the following formula (5).

Here, δσ _i σ _j is the correlation between the particle i and the particle j, J _ij is the interaction between the particle i and the particle j, and the sum of these products is the neighborhood region <i, j It is defined by>. Now, if there are N particles, σ = (σ ₁ ,..., Σ _N ) represents the state of this system, and the probability of becoming a certain state σ _h is expressed by the following equation (6) ). However, z is a constant for normalization.

  FIG. 4 is a schematic diagram for explaining the states of ferromagnetism, paramagnetism, and superparamagnetism. When the temperature T of the system is changed, a phase transition occurs. When T is small, the ferromagnetism shown in FIG. 4A is shown, and when T is large, it is shown in FIG. 4B. It becomes paramagnetism. In ferromagnetism, the spin converges in one direction, and in paramagnetism, the spins fall apart and have no order, but the super-paramagnetism shown in Fig. 4 (c) may appear between them. . This is a state in which the magnetic domain is small and the magnetization is fluctuating thermally. That is, a particle becomes a cluster, and a plurality of clusters exist. For this reason, in the present embodiment, the motion of the image is clustered by determining the temperature at which superparamagnetism is obtained by regarding the pixel as a particle.

Further, when there is a motion in the image, an optical flow occurs. Now, let the luminance of the image be I (x, y, t). The assumption is made that the luminance before and after the movement of a certain pixel does not change between the two images. I (x, y, t) and x, y, a material obtained by partially differentiating _{t bf = (f x, f} y, f t), when the motion vector of the pixel bu = (u, v) and, below The constraint equation shown in equation (7) is obtained.

This is called BCC (Brightness Constant Constraint) and indicates that the motion vector rides on a straight line represented by bf on the image. However, u and v cannot be determined uniquely from this constraint equation alone. . For this reason, in the present embodiment, a method of explicitly using cluster information is used. Assuming that the motion of the temporary cluster to which the pixel i belongs is BU _i = (U _i , V _i ), if the pixel i belongs to this cluster, BU _i is the constraint of the above equation (7). Should meet. Therefore, the motion similarity d _ij between the pixel i and the pixel j shown in the above equation (1) is introduced as an index of the motion similarity between the pixel i and the pixel j. In the equation (1), d _i represents the distance between the cluster motion vector BU _i and the constrained straight line represented by the partial differentiation.

Thus, in the Potts model, the interaction J _ij between the particles expressed by the above formula (2) plays an important role, and in this embodiment, the pixel belongs to the BCC of the pixel. By determining that the motion vectors of the clusters match and determining the interaction J _ij , the interaction J depends on the motion similarity d _ij between the pixel i and the pixel j represented by the above equation (1). _ij is defined.

  Next, image analysis processing by the image analysis apparatus configured as described above will be described. FIG. 5 is a flowchart for explaining image analysis processing by the image analysis apparatus shown in FIG.

  First, when the user inputs an image data acquisition command, in step S <b> 1, the image acquisition unit 11 acquires image data including, for example, 480 × 640 pixels and outputs the image data to the preprocessing unit 12. Next, in step S <b> 2, the preprocessing unit 12 performs a smoothing process on the entire image data using a Gaussian filter, reduces the resolution to 1/8, and outputs the result to the clustering unit 13.

Next, in step S3, the clustering section 13, partial differential bf = the brightness of each pixel _{(f x, f y, f} t) is calculated. Next, in step S4, the clustering unit 13 calculates the Curie temperature T _C and sets a temperature sweep range based on the calculated T _C. Next, in step S5, the clustering unit 13 sets the temperature within the set temperature sweep range.

In the Potts model, the behavior of the system is determined by the temperature T and the interaction J _ij . In the present embodiment, J _ij shown in the above equation (2) is used, and the temperature T is swept from a low temperature to a high temperature to find an appropriate temperature. In general, the phase transition from ferromagnetic to paramagnetic is to occur at the Curie temperature T _C, the Curie temperature T _C is approximated by the following equation (8).

It is considered that even superparamagnetic appearing in the vicinity this, in this embodiment, the sweep range of the temperature T 0 <T <4T _C. The magnetic susceptibility χ at each temperature is used as an index of how the phase transition has changed.

  Next, in step S <b> 6, the clustering unit 13 performs a clustering process using a Swendsen-Wang (SW) algorithm. FIG. 6 is a flowchart for explaining the clustering process in step S6 shown in FIG.

First, in step S21, the clustering unit 13 regards each pixel in the image as a particle, and randomly assigns spin states (1,..., Q) to σ _i (i = 1,..., N). Next, in step S22, the clustering unit 13 sets a cluster for each pixel, and sets a random value for the motion vector of the cluster. Therefore, each pixel is a cluster having 1 pixel, and the total number of clusters is N.

Next, in step S23, the clustering unit 13 uses Equation (2) using the motion similarity d _ij between the pixel i and the pixel j shown in Equation (1) in the nearest eight neighboring regions of all the pixels. The interaction J _ij between the particle i and the particle j shown in FIG. Next, in step S24, the clustering unit 13 cuts or joins the links with the neighboring pixels according to the probabilities shown in the above formulas (3) and (4). However, the coupling is limited only to σ _i = σ _j .

  Next, in step S25, the clustering unit 13 sets a set of pixels having combined links as a cluster. Next, in step S26, the clustering unit 13 calculates the motion vector BU of each cluster by combining the constraint equations (7) for the pixels included in the cluster. When the number of pixels included in the cluster is small, or when the solution becomes indefinite even if the constraint equation (7) is coupled, a random value is stored in BU. Next, in step S27, the clustering unit 13 randomly assigns a spin value to each pixel. However, the same spin value is assigned to pixels belonging to the same cluster.

  Next, in step S28, the clustering unit 13 determines whether or not the processing of steps S23 to S27 has been performed in advance for the number of iterations of the Monte Carlo method. If the number of iterations has not been reached, the processing of steps S23 to S27 is performed. When the number of iterations is reached, the iteration process is terminated and the process proceeds to step S7 shown in FIG.

Referring to FIG. 5 again, in step S7, the clustering unit 13 calculates the magnetic susceptibility and the like as follows. First, the clustering unit 13 sets c _ij to 1 when the pixel i and the pixel j belong to the same cluster, sets c _ij to 0 in other cases, and sets the average of all steps of the Monte Carlo method as <c> Calculate _ij >. Next, the clustering unit 13 uses this <c _ij > to calculate the probability <δ _ij > = ((q−1) <c _ij > +1) / q that the pixel i and the pixel j belong to the same cluster. To do. Next, the clustering unit 13 combines the links of pixels that are <δ _ij >> θ (where θ is a threshold for determining a cluster), and sets a set of pixels having the connection as a cluster. This is the result of clustering at the temperature T set in step S5. Finally, the clustering unit 13 calculates the magnetic susceptibility χ = N (<m ² > − <m> ² ) / T using the dispersion of the magnetization m. The magnetization m is calculated as m = (qN _max −N) / (N (q−1)) in each Monte Carlo method step. However, N _max = max (N ₁ ,..., N _q ), which is the maximum number of pixels taking each spin value.

  Next, in step S8, the clustering unit 13 determines whether or not the processing in steps S5 to S7 has been executed for the sweep range set in step S4 above, and the processing for all the sweep ranges has been completed. If not, the temperature T is changed by a predetermined value in step S5 and the processes in steps S6 to S7 are repeated. If the processes for all the sweep ranges are completed, the process proceeds to step S9.

As described above, when the process is repeated while sweeping the temperature, the magnetic susceptibility chi, when the temperature is low, i.e. approximately 0 in the state of ferromagnetic, then rises sharply in the vicinity Curie temperature T _C. When the temperature is further increased, the magnetic susceptibility χ decreases, and a range that does not change for a while appears. This is a superparamagnetic state, which is a temperature suitable for clustering. After that, it suddenly drops to 0 and becomes paramagnetic.

  Accordingly, in step S9, the clustering unit 13 sets the temperature at which the magnetic susceptibility χ changes as described above to take a value within a predetermined range, for example, the intermediate temperature among the temperatures at which the change amount of the magnetic susceptibility χ falls within a predetermined value. Is determined as the temperature at which the particles become superparamagnetic. Next, in step S <b> 10, the clustering unit 13 outputs the clustering result of the image motion at the determined temperature to the motion extracting unit 14.

  Next, in step S <b> 11, the motion extraction unit 14 extracts a motion cluster in the image to be processed based on the clustering result of the clustering unit 13. Finally, in step S12, the image processing unit 15 executes predetermined image processing based on the extracted cluster to extract image motion or classify image motion, and the image analysis processing ends.

Figure 7 is a diagram showing an example of an image representing the front and rear motion, FIG. 8 is a diagram showing the f _x, f _y, f _t between the images shown in FIG. Using two images shown in FIG. 7 result of the image analysis processing for clustering movement of the image, f _x between the images shown in (a) ~ (c) of FIG. 8, f _y, f _t was gotten. Here, the processing conditions, but the original image shown in FIG. 7 is a 480 × 640 pixels, f _x, f _y, In calculating f _t, and first pre-smoothing the entire image Gaussian filter The resolution was set to 1/8, the neighborhood of 8 pixels was used as the neighborhood region, and the spin value was set to q = 11. In addition, the number of iterations of the Monte Carlo method was 4000, and the parameter θ for determining a cluster was 0.5.

  FIG. 9 is a diagram showing the relationship between the magnetic susceptibility χ obtained by the image analysis processing on the image shown in FIG. 7 and the temperature T, and FIG. 10 shows an example of the clustering result when the temperature T = 0.25. FIG. As shown in FIG. 9, it was found that the change in magnetic susceptibility χ decreased near T = 0.25, indicating a superparamagnetic state. The clustering result at this time is shown in FIG. 10, and it can be seen that the head and the torso move almost the same, and the right arm moves differently, and the motion of the image is extracted with high accuracy. I was able to.

Through the above processing, in the present embodiment, each pixel of image data is treated as a particle using the Potts model, and the motion of the image in a state where the particle becomes superparamagnetic is clustered by a statistical method using the Sendsen-Wang algorithm. Then, a motion cluster in the image is extracted based on the clustering result, and in this clustering process, an expression defined using the motion similarity d _ij between the pixel i and the pixel j shown in the expression (1) above. Since the motion of the image is clustered using the interaction J _ij between the particle i and the particle j shown in (2), the superparamagnetic state appears clearly, and the motion of the image at this time is clustered. Movement can be extracted with high accuracy. Moreover, since global movements can be clustered simply by defining local interactions as described above, prior knowledge can be reduced.

  In the above description, the example in which the image analysis processing according to the present invention is performed using software has been described. However, part or all of the image analysis processing may be configured using dedicated hardware. Processing can be executed.

It is a block diagram which shows the structure of the image analysis apparatus by one embodiment of this invention. It is a main functional block diagram of the image analysis apparatus shown in FIG. It is a schematic diagram of an Ising model and a Potts model. It is a schematic diagram for demonstrating the state of ferromagnetism, paramagnetism, and superparamagnetism. 3 is a flowchart for explaining image analysis processing by the image analysis apparatus shown in FIG. 2. It is a flowchart for demonstrating the clustering process of step S6 shown in FIG. It is a figure which shows an example of the image showing the back and front of a motion. F _x between the image shown in FIG. 7 is a diagram showing the f _y, f _t. It is a figure which shows the relationship between the magnetic susceptibility χ obtained by the image analysis process with respect to the image shown in FIG. It is a figure which shows an example of the clustering result in case temperature T = 0.25.

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 Image acquisition part 12 Pre-processing part 13 Clustering part 14 Motion extraction part 15 Image processing part

Claims (7)

Acquisition means for acquiring image data to be processed;
Each pixel of the image data acquired by the acquisition means is treated as a particle using a spin model of a magnetic material in which the spin state is expanded to multiple values, and the motion of the image in a state in which the particle is superparamagnetic by a statistical method Clustering means for clustering;
Extracting means for extracting a motion cluster in the image based on the clustering result by the clustering means;
The image analysis apparatus characterized in that the clustering means clusters the movement of an image using an interaction between pixels defined by explicitly using the movement of the cluster. Cluster The clustering means, the brightness of the image I (x, y, t) , I (x, y, t) partially differentiated bf = of _{(f x, f y, f} t), and the pixel i belongs When the motion vector of BU _i = (U _i , V _i ) is used, particles represented by the following formula (2) are used by using the motion similarity d _ij between the pixel i and the pixel j represented by the following formula (1). calculating the interaction J _ij between i and the particle j (where K is the average number of neighboring pixels, a is the average of d _ij ), and clustering the image motion using the calculated interaction J _ij The image analysis apparatus according to claim 1.
  The image analysis apparatus according to claim 2, wherein the clustering means clusters image motion using a Monte Carlo method. The clustering means assigns a spin value to each pixel at random using the Swendsen-Wang algorithm, and cuts or joins the link between the target pixel and the neighboring pixel according to the probability expressed by the following formulas (3) and (4). A group of pixels having linked links is used as a cluster, a motion vector BU _i in each cluster is calculated, and a process of assigning a spin value randomly to each pixel so that the same spin value is assigned to a pixel belonging to the same cluster a predetermined number of times The image analysis apparatus according to claim 3, wherein the motion of the image is clustered by performing iterative processing only.
  The clustering means determines the temperature at which the magnetic susceptibility calculated using the interaction between pixels takes a value within a predetermined range when the temperature is swept, as the temperature at which the particles are superparamagnetic. The image analysis apparatus according to claim 1, wherein the movement of the image at the determined temperature is clustered. Acquisition means for acquiring image data to be processed;
Each pixel of the image data acquired by the acquisition means is treated as a particle using a spin model of a magnetic material in which the spin state is expanded to multiple values, and the motion of the image in a state in which the particle is superparamagnetic by a statistical method Clustering means for clustering;
Causing the computer to function as an extracting means for extracting a cluster of motion in an image based on the clustering result by the clustering means;
The clustering means clusters an image motion using an interaction between pixels defined by explicitly using a cluster motion. An image analysis method for analyzing the movement of an image using a computer that functions as an acquisition unit, a clustering unit, and an extraction unit,
A first step in which the acquisition means acquires image data to be processed;
The clustering means treats each pixel of the image data acquired by the acquisition means as a particle using a spin model of a magnetic material in which the spin state is expanded to multiple values, and the image motion in a state where the particle is superparamagnetic A second step of clustering with statistical methods;
The extraction means includes a third step of extracting a motion cluster in the image based on a clustering result by the clustering means;
The image analysis method according to claim 2, wherein the second step includes a step of clustering the motion of the image using the interaction between the pixels defined by explicitly using the motion of the cluster.