Image processing device, image recognition device, image processing program, and image recognition program
An image processing device has a function for plotting a luminance gradient co-occurrence pair of an image on a feature plane and applying an EM algorithm to form a GMM. The device learns a pedestrian image and creates a GMM, subsequently learns a background image and creates a GMM, and calculates a difference between the two and generates a GMM for relearning based on the calculation. The device plots a sample that conforms to the GMM for relearning on the feature plane by applying an inverse function theorem. The device forms a GMM that represents the distribution of samples at a designated mixed number and thereby forms a standard GMM that serves as a standard for image recognition. When this mixed number is set to less than a mixed number designated earlier, the dimensions with which an image is analyzed are reduced, making it possible to reduce calculation costs.
1. An image processing device comprising:
at least one processor programmed to:
input an image;
acquire a combination of two pixels from the inputted image;
acquire a co-occurrence in luminance gradient directions of the two pixels related to the acquired combination;
acquire a distribution of an occurrence frequency of the co-occurrence that is acquired while changing the acquired combination of the two pixels;
generate a probability density function (p(x1θ)) corresponding to the acquired distribution of the occurrence frequency, the probability density function (p(x1θ)) being a Gaussian Mixture Model (GMM), which is expressed by:
p ( x 1θ)=Σ K j=1 p ( x | z=j ,θ)=Σ K j=1 π j N ( x |μ j Σ j )
where K is a number of mixture coefficients, j=1, 2, . . . , K, x is a vector quantity expressing a distribution of co-occurrence corresponding points, θ is a vector quantity expressing a population parameter, π j is a mixture coefficient indicating a probability for selecting a j-th Gaussian distribution, μ j and Σ j respectively indicate an average value of j-th Gaussian distribution and a variance-covariance matrix, and z is a potential parameter for calculating an Expectation maximization (EM) and an algorithm and load ratio γ, and z1, z2, . . . , zK are used in correspondence with the K Gaussian distributions to be mixed, the load ratio γ being a probability that a certain co-occurrence corresponding point of the distribution of co-occurrence corresponding points is generated from the j-th Gaussian distribution;
output the generated probability density function in accordance with a parameter that includes π j and θ to define the generated probability density function;
input a plurality of recognizing object images including an inputted recognizing object;
generate a positive Gaussian Mixture Model composed of a recognizing object probability density function fp(x) based on the plurality of recognizing object images;
input a plurality of non-recognizing object images, which do not include the inputted recognizing object, and generate a negative Gaussian Mixture Model composed of a non-recognizing object probability density function fn(x) according to the plurality of non-recognizing object images;
generate metric information to adopt a metric to a correlation between the generated recognizing object probability density function and the generated non-recognizing object probability density function;
generate a Gaussian Mixture Model composed of a reference probability density function g′(x) based on the generated metric information, the generated reference probability density function being a reference for executing image recognition of the recognizing object using the generated recognizing object probability density function; and
output the generated reference probability density function in accordance with the parameter for defining the generated reference probability density function.
2. The image processing device according to claim 1 , wherein the at least one processor is programmed to generate the metric information in which the metric of the correlation with fn(x) viewed from fp(x) and the metric of the correlation with fp(x) viewed from fn(x) are equal to each other.
3. The image processing device according to claim 2 , wherein the at least one processor is programed to:
generate a plurality of samples in accordance with the generated metric information, and
generate the reference probability density function based on a distribution of an occurrence frequency of the generated sample.
4. The image processing device according to claim 3 , wherein
the recognizing object probability density function and the non-recognizing object probability density function are generated respectively by mixing a predetermined number of basis functions, and
the at least one processor is programmed to mix a number of the basis functions that is smaller than the predetermined number to generate the reference probability density function.
5. The image processing device according to claim 2 , wherein
the recognizing object probability density function and the non-recognizing object probability density function are generated respectively by mixing a predetermined number of basis functions, and
the at least one processor is programmed to mix a number of the basis functions that is smaller than the predetermined number to generate the reference probability density function.
6. The image processing device according to claim 1 , wherein the at least one processor is programed to:
generate a plurality of samples in accordance with the generated metric information, and
generate the reference probability density function based on a distribution of an occurrence frequency of the generated sample.
7. The image processing device according to claim 6 , wherein
the recognizing object probability density function and the non-recognizing object probability density function are generated respectively by mixing a predetermined number of basis functions, and
the at least one processor is programmed to mix a number of the basis functions that is smaller than the predetermined number to generate the reference probability density function.
8. The image processing device according to claim 1 , wherein
the recognizing object probability density function and the non-recognizing object probability density function are generated respectively by mixing a predetermined number of basis functions, and
the at least one processor is programmed to mix a number of the basis functions that is smaller than the predetermined number to generate the reference probability density function.
9. The image processing device according to claim 1 , wherein the at least one processor is programmed to:
convert a resolution of the input image, and
set a pixel of the image with the converted resolution to be an object of the combination.
10. The image processing device according to claim 1 , wherein the metric information is a differential probability density function using a difference between the recognizing object probability density function and the non-recognizing object probability density function.
11. The image processing device according to claim 1 , wherein the metric information is a ratio probability density function using a ratio between the recognizing object probability density function and the non-recognizing object probability density function.
12. The image processing device according to claim 11 , wherein the ratio probability density function further uses a difference between the recognizing object probability density function and the non-recognizing object probability density function.
13. An image recognition device comprising:
a computer programmed to:
acquire the reference probability density function output from the image processing device according to claim 1 ;
acquire an object image which is an image recognition object;
acquire a feature amount of the acquired object image using the acquired reference probability density function; and
determine whether the object image includes the image recognition object using the acquired feature amount.
14. The image recognition device according to claim 13 , wherein the computer is programmed to:
acquire a distribution of an occurrence frequency of co-occurrence in gradient directions in the object image, and
acquire the feature amount of the object image based on a probability that the acquired distribution of the occurrence frequency is generated from the acquired reference probability density function.
15. A non-transitory computer readable storage medium storing computer-executable instructions of an image recognition program that, when executed by a computer, causes the computer to perform processes comprising:
acquiring the reference probability density function output from the image processing device according to claim 1 ;
acquiring an object image which is an image recognition object;
acquiring a feature amount of the acquired object image using the acquired reference probability density function; and
determining whether the object image includes the image recognition object using the acquired feature amount.
16. A non-transitory computer readable storage medium storing computer-executable instructions of an image processing program that, when executed by a computer, causes the computer to perform processes comprising:
inputting an image;
acquiring a combination of two pixels from the inputted image;
acquiring a co-occurrence in luminance gradient directions of the two pixels related to the acquired combination;
acquiring a distribution of an occurrence frequency of the co-occurrence that is acquired while changing the acquired combination of the two pixels;
generating a probability density function (p(x1θ)) corresponding to the acquired distribution of the occurrence frequency, the probability density function (p(x1θ)) being a Gaussian Mixture Model (GMM), which is expressed by:
p ( x 1θ)=Σ K j=1 p ( x | z=j ,θ)=Σ K j=1 π j N ( x |μ j Σ j )
where K is a number of mixture coefficients, j=1, 2, . . . , K, x is a vector quantity expressing a distribution of co-occurrence corresponding points, θ is a vector quantity expressing a population parameter, π j is a mixture coefficient indicating a probability for selecting a j-th Gaussian distribution, μ j and Σ j respectively indicate an average value of j-th Gaussian distribution and a variance-covariance matrix, and z is a potential parameter for calculating an Expectation maximization (EM) and an algorithm and load ratio γ, and z1, z2, . . . , zK are used in correspondence with the K Gaussian distributions to be mixed, the load ratio γ being a probability that a certain co-occurrence corresponding point of the distribution of co-occurrence corresponding points is generated from the j-th Gaussian distribution;
outputting the generated probability density function in accordance with a parameter that includes π j and θ to define the generated probability density function;
inputting a plurality of recognizing object images including an inputted recognizing object;
generating a positive Gaussian Mixture Model composed of a recognizing object probability density function fp(x) based on the plurality of recognizing object images;
inputting a plurality of non-recognizing object images, which do not include the inputted recognizing object, and generating a negative Gaussian Mixture Model composed of a non-recognizing object probability density function fn(x) according to the plurality of non-recognizing object images;
generating metric information to adopt a metric to a correlation between the generated recognizing object probability density function and the generated non-recognizing object probability density function;
generating a Gaussian Mixture Model composed of a reference probability density function g′(x) based on the generated metric information, the generated reference probability density function being a reference for executing image recognition of the recognizing object using the generated recognizing object probability density function; and
outputting the generated reference probability density function in accordance with the parameter for defining the generated reference probability density function.