Visual feature extraction methods and systems
View Patent ↗Systems and methods for image processing are disclosed. For example, a system for processing images includes voxel-ranking circuitry that ranks individual voxels using voxel data derived from physiological responses resulting from exposure to a plurality of corresponding baseline stimulus images, and voxel selection circuitry that selects a portion of the set of ranked voxels based on their rank.
1. A system for processing images, comprising:
voxel-ranking circuitry that ranks individual voxels using voxel data derived from physiological responses resulting from exposure to a plurality of corresponding baseline stimulus images;
voxel selection circuitry that selects a portion of the set of ranked voxels based on their rank; and
filter circuitry that creates two or more filters by determining which voxels will activate maximally in response to a feature presented in one or more stimulus images, then isolating each filter, the isolation of each filter based on the following equation:
φ i =∫r i ( t ) Sdt=∫r i ( t )Σ i φ i *r i ( t ) dt ,
where φ i represents an average filter response and expresses that a signal S is constructed from a weighted linear sum of neural firing rates, r i (t) represents variance-normalized average activity in a voxel at time step t, and S represents the stimulus image.
2. The system of claim 1 , wherein:
the voxel-ranking circuitry ranks voxels based upon which voxels have a lowest variance.
3. The system of claim 1 , wherein:
the voxel-ranking circuitry ranks voxels based upon information redundancy of a voxel as compared to other voxels.
4. The system of claim 1 , wherein:
the voxel-ranking circuitry ranks voxels based upon information entropy of individual voxels.
5. The system of claim 4 , further comprising:
component analysis circuitry that reduces each of the stimulus images used to create the voxel data to a set of vectors.
6. The system of claim 5 , wherein the component analysis circuitry reduces each of the stimulus images using a Principal Components Analysis (PCA) technique so as to identify eigenvectors of an image.
7. The system of claim 6 , further comprising:
weighting circuitry that weights the eigenvectors so as to form a loading matrix.
8. The system of claim 7 , wherein the voxel ranking circuitry uses a Bayesian regression on the voxel data along with the loading matrix to rank voxels.
9. The system of claim 5 , wherein the component analysis circuitry reduces each of the stimulus images using an Independent Component Analysis (ICA) technique so as to identify vectors of an image.
10. The system of claim 1 , wherein the voxel ranking circuitry uses a Bayesian regression on the voxel data with stimulus image data to rank voxels.
11. The system of claim 10 , wherein Bayesian regression conforms to the following equation:
A i ( t )=Σγ im Σβ mj S j ( t ),
where A j represents a projection of the i th component onto a stimulus image presented at time t, S j represents the activity of voxel j at time t, γ ij is a binary selection vector, and β ij is a matrix of regression coefficients.
12. A method for processing images, comprising:
ranking individual voxels using voxel data derived from physiological responses resulting from exposure to a plurality of corresponding baseline stimulus images;
selecting a portion of ranked voxels based on their rank, and
creating two or more filters by determining which voxels will activate maximally in response to a feature presented in one or more stimulus images, then isolating each filter, the isolation of each filter based on the following equation:
φ i =∫r i ( t ) Sdt=∫r i ( t )Σ i φ i *r i ( t ) dt ,
where φ i represents an average filter response and expresses that a signal S is constructed from a weighted linear sum of neural firing rates, r i (t) represents variance-normalized average activity in a voxel at time step t, and S represents the stimulus image.
13. The method of claim 12 , wherein:
ranking individual voxels is based upon which voxels have a lowest variance and/or information redundancy as compared to other voxels.
14. The method of claim 12 , wherein:
ranking individual voxels is based upon information entropy of individual voxels.
15. The method of claim 14 , further comprising:
reducing each of the stimulus images to a set of vectors.
16. The method of claim 15 , wherein reducing each of the stimulus images is based on at least one of a Principal Components Analysis (PCA) technique or an Independent Component Analysis (ICA) technique so as to identify vectors of an image.
17. The method of claim 16 , further comprising:
weighting the vectors so as to form a loading matrix.
18. The method of claim 17 , wherein the Bayesian regression technique conforms to the following equation:
A i ( t )=Σγ im Σβ mj S j ( t ),
where A i represents a projection of the i th component onto a stimulus image presented at time t, S j represents the activity of voxel j at time t, γ ij is a binary selection vector, and β ij is a matrix of regression coefficients.
19. The method of claim 12 , wherein ranking individual voxels includes applying a Bayesian regression technique on the voxel data with stimulus image data.