A transfer learning method from macro-expression to micro-expression
A transfer learning and micro-expression technology, applied in the field of computer vision and pattern recognition, can solve the problems of limited number of micro-expression database samples, unsatisfactory recognition effect, limited binding force, etc.
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0080] A transfer learning method from macro-expression to micro-expression, such as figure 1 shown, including the following steps:
[0081] (1) Extract macro-expression features and micro-expression features respectively: extract LBP features from macro-expression to form macro-expression feature representation matrix X a , micro-expression extraction LBP-TOP features to form a micro-expression feature representation matrix X b , and take the average of the three orthogonal planes of the LBP-TOP feature. The original LBP-TOP feature is the LBP feature extracted from the three orthogonal planes of XY, XT, and YT, forming a 3*59-dimensional feature. Here, for two The dimensions of expression are unified, and the LBP features on the three orthogonal planes are accumulated and averaged to form a 59-dimensional LBP-TOP feature, forming a micro-expression representation feature that is unified with the macro-expression dimension;
[0082] (2) Project the micro-expression feature ...
Embodiment 2
[0087] According to a method of transfer learning from macro-expressions to micro-expressions described in Embodiment 1, the difference is that,
[0088] In step (2), the objective function is as shown in formula (I):
[0089]
[0090] In formula (I), α, β, γ are balance coefficients, α>0, β>0, γ>0;
[0091] x a is the macro-expression feature matrix, n a Represents the number of macro expression samples, d represents the feature dimension;
[0092] x b is the micro-expression feature matrix, n b Represents the number of dimension expression samples, and d represents the feature dimension;
[0093] P∈R d×np , P is the projection matrix, np is the subspace dimension;
[0094] V is the reconstruction coefficient matrix (equivalent to the secondary projection matrix);
[0095] E is an error matrix, which stores the respective characteristic components in the macro-expression and micro-expression, and the characteristic components include their unique characteri...
Embodiment 3
[0114] According to a method of transfer learning from macro-expression to micro-expression described in embodiment 1 or 2, the difference is that
[0115] In step (3), the projection matrix P is initialized by NMF, and the target projection matrix is iteratively optimized by solving the objective function, and the target projection matrix P is output when the convergence condition is met. 1 , including the following steps:
[0116] E. Convert the objective function into an approximate form for solution: such as figure 2 said;
[0117] The objective function is shown in formula (Ⅵ):
[0118]
[0119] Considering various constraints, the objective function is non-convex, and the objective function is transformed into an approximate form, as shown in formula (VII):
[0120]
[0121] In formula (VII), Y is by the eigenvector that WY=ΛDY problem solves as the matrix that row vector forms, Λ is to form diagonal matrix as diagonal element by corresponding eigenvalue;
...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More 


