초록 ▼

A method for performing shape localization in an image includes deriving a model shape from a database of a plurality of sample shapes. The model shape is defined by a set of landmarks. The method further includes deriving a texture likelihood model of present sub-patches of the set of landmarks def

A method for performing shape localization in an image includes deriving a model shape from a database of a plurality of sample shapes. The model shape is defined by a set of landmarks. The method further includes deriving a texture likelihood model of present sub-patches of the set of landmarks defining the model shape in the image, and proposing a new set of landmarks that approximates a true location of features of the shape based on a sample proposal model of the present sub-patches. A CONDENSATION algorithm is used to derive the texture likelihood model and the proposed new set of landmarks.

대표청구항 ▼

The invention claimed is: 1. A method for performing face shape localization in an image, comprising: deriving a model face shape from a database of a plurality of sample face shapes, said model face shape being defined by a set of landmarks; deriving a texture likelihood model of present sub-patch

The invention claimed is: 1. A method for performing face shape localization in an image, comprising: deriving a model face shape from a database of a plurality of sample face shapes, said model face shape being defined by a set of landmarks; deriving a texture likelihood model of present sub-patches of said set of landmarks defining said model face shape in the image; and proposing a new set of landmarks that approximates a true location of features of the face shape based on a sample proposal model of said present sub-patches; wherein said deriving said texture likelihood model and said proposing said new set of landmarks are conducted using a CONDENSATION algorithm; and said model face shape is derived from a prior probabilistic distribution of a predefined model p(m), said texture likelihood model of said present sub-patches is derived from a local texture likelihood distribution model p(I|m), and said sample proposal model is derived based on a texture likelihood model of subsequent sub-patches of a set of landmarks in the image at proposed locations in a vicinity of said present sub-patches of said present set of landmarks. 2. The method as defined in claim 1, wherein the a prior probabilistic distribution of a predefined model p(m) describes a 2-dimensional shape represented as a vector, description="In-line Formulae" end="lead"S=(x1, x2, . . . , xK, y1, y2, . . . , yK)T description="In-line Formulae" end="tail" of length 2K, where K is a number of landmarks that define said face shape. 3. The method as defined in claim 2, wherein said face shape is modeled as a vector description="In-line Formulae" end="lead"S= S+Uw,description="In-line Formulae" end="tail" when aligned with at least one of a plurality of manually labeled shape images by taking a first k principal components, where S is a mean shape, and U2K��k is an eigenvector matrix, and wk��1 is a parameter vector that define said shape S. 4. The method as defined in claim 3, wherein the a prior probabilistic distribution of a predefined model p(m) is obtained by learning a mixture of Gaussian model after projecting said shape vector S in the k dimensional active shape model (ASM) eigenspace. 5. The method as detined in claim 3 wherein, said shape vector S is rearranged as where ({circumflex over (��)}) denotes the rearrangement operation of elements in the shape vector. 6. The method as defined in claim 5 wherein, said shape vector S is denoted as where S is a scaling factor, θ is an angle of rotation, and is the translation in the image, wherein the landmarks of the face shape in image is represented as parameter model description="In-line Formulae" end="lead"m=(s, θ, T, w).description="In-line Formulae" end="tail" 7. The method as defined in claim 1, wherein the local texture likelihood distribution model p(I|m) is defined as supposing the texture of each landmark is independent, and where Γj denotes a sub-patch of landmark j defining a two-dimensional shape; wherein a texture likelihood p(Γj) of landmark i is independently learned as a Mixture of Gaussian model of sub-patches of each landmark cropped from a plurality of manually labeled training images in the database projected into their customized feature subspaces. 8. The method as defined in claim 7, wherein description="In-line Formulae" end="lead"I1={Γ1(i), Γ2(i), . . . , ΓK(i)}description="In-line Formulae" end="tail" is a collection of local observation of shape features in image at interation i, and by regarding a predetined model {p1, p2, . . . , pK} as a landmark set, and assuming independence of a p(m/Ii-1) of each of said landmark p(pj(i)|Γj(i)); said p(m/Ii-1) of each landmark is formulated as where Γ(x, y) means a subpatch centered at (x, y); and description="In-line Formulae" end="lead"p(pj=(x, y)|Γ(x, y))˜p(Γ(x, y)|pj=(x, y))p(pj=(x, y))=p(Γ(x, y)j)p(pj=(x, y)),description="In-line Formulae" end="tail" where p(Γ(x, y)j) is a texture likelihood of landmark j at location (x, y), and p(pj=(x, y)) is modeled as a uniform distribution in the image. 9. The method as defined in claim 8, wherein the formula for the proposal model of each landmark is converted to model parameter space expressed by an equation, description="In-line Formulae" end="lead"Δm(i)=(Δs(i), Δθ(i), ΔT(i), Δw(i)), and m(i+1)=m(i)+aΔm(i) for some 01.description="In-line Formulae" end="tail" 10. The method as defined in claim 9, wherein said model parameter space equation is obtained from a new model sample proposed as description="In-line Formulae" end="lead"{p1(i), p2(i), . . . , pk(i)},description="In-line Formulae" end="tail" by supposing the rotation angle is very small, by taking derivative of Xi, Yi with respect to θ, T, and w', 11. The method as defined in claim 1, wherein the CONDENSATION algorithm is performed separately on at least one reduced-size of the image prior to performing the CONDENSAILON algorithm on a full-size of the image. 12. The method as defined in claim 11, wherein the CONDENSATION algorithm is performed separately tbr a plurality of image resolutions starting from a low image resolution to a high image resolution. 13. The method as defined in claim 12, wherein the number of landmarks defining said model face shape is increased from the low image resolution to the high image resolution. 14. The method as defined in claim 13, wherein dimension of shape eigen-space arc also increased from the low image resolution to the high image resolution. 15. method as defined in claim 11, wherein the CONDENSATION algorithm is performed hierarchically on a plurality of image resolutions starting from a low image resolution to a high image resolution. 16. Method for performing a face localization in an image based on a Bayesian rule, comprising: deriving a predefined face shape model m; employing conditional density propagation (CONDENSATION) algorithm to locate a face shape in the image using a prior probabilistic distribution of a model p(m) based on said predefined face shape model m, and a local texture likelihood distribution given said predefined face shape model with specific model parameters p(I|m). 17. The method as defined in claim 16, wherein the CONDENSATION algorithm is performed separately on at least one reduced-size of the image prior to performing the CONDENSATION algorithm on a full-size of the image. 18. The method as defined in claim 16, wherein the a prior probabilistic distribution of a predefined model p(m) describes a 2-dimensional face shape represented as a vector, description="In-line Formulae" end="lead"S=(x1, x2, . . . , xK, y1, y2, . . . , yK)T description="In-line Formulae" end="tail" of length 2K, where K is a number of landmarks that define said face shape. 19. The method as defined in claim 18, wherein the face shape is modeled as a vector description="In-line Formulae" end="lead"S= S+Uw,description="In-line Formulae" end="tail" when aligned with at least one of a plurality of manually labeled face images by taking a first k principal components, where S is a mean face shape, and U2K��k is an eigenvector matrix, and wk��1 is a parameter vector that define said face shape S. 20. The method as defined in claim 19, wherein the a prior probabilistic distribution of a predefined model p(m) is obtained by learning a mixture of Gaussian model after projecting said face vector S in the k dimensional active shape model (ASM) eigenspace. 21. The method as defined in claim 20 wherein, said face shape vector S is rearranged as where ({circumflex over (��)}) denotes the rearrangement operation of shape vector. 22. The method as defined in claim 21 wherein, said shape vector S is denoted as where s is a scaling factor, θ is an angle of rotation, and is the translation in the image, wherein the landmarks of the face in image is represented as parameter model description="In-line Formulae" end="lead"m=(s, 0, T, w).description="In-line Formulae" end="tail" 23. The method as defined in claim 16, wherein the local texture likelihood distribution model p(I|m) is defined as supposing the texture of each landmark is independent, and where Γj denotes a sub-patch of landmark j defining a two-dimensional face shape; wherein a texture likelihood p(Γj) of landmark i is independently learned as a Mixture of Gaussian model of sub-patches of each landmark cropped from a plurality of manually labeled training images in the database projected into their customized feature subspaces. 24. The method as defined in claim 23, wherein description="In-line Formulae" end="lead"Ii={Γ1(i), Γ2(i), . . . , ΓK(i)}description="In-line Formulae" end="tail" is a collection of local observation of facial features in image at interation i, and by regarding a predefined model {p1, p2, . . . , pK} as a landmark set, and assuming independence of a p(mi/Ii-1) of each of said landmark p(pj(i)|Γj(i)); said p(mi/Ii-1) of each landmark is formulated as where Γ(x, y) means a subpatch centered at (x, y); and description="In-line Formulae" end="lead"p(pj=(x, y)|Γ(x, y))˜p(Γ(x, y)|pj=(x, y))p(pj=(x, y))=p(Γ(x, y)j)p(pj=(x, y))description="In-line Formulae" end="tail" where p(Γ(x, y)j) is a texture likelihood of landmark j at location (x, y), and p(pj=(x, y)) is modeled as a uniform distribution in the image. 25. The method as defined in claim 24, wherein the formula for the proposal model of each landmark is converted to model parameter space expressed by an equation, description="In-line Formulae" end="lead"Δm(i)=(Δs(i), Δθ(i),ΔT(i), Δw(i)), and m(i+1)32 m(i)+aΔm(i) for some 01.description="In-line Formulae" end="tail" 26. The method as defined in claim 25, wherein said model parameter space equation is obtained from a new model sample proposed as description="In-line Formulae" end="lead"{p1(i), p2(i), . . . , pk(i)},description="In-line Formulae" end="tail" by supposing the rotation angle is very small, by taking derivative of Xi, Yi with respect to θ, T, and w',