Object recognition using a convolutional neural network trained by principal component analysis and repeated spectral clustering
A method of object recognition trains a convolutional neural network (CNN) with a set of training images, then classifies an image of an object using the trained CNN. A first layer of the CNN is trained by generating a set of first convolutional filters from eigenvectors produced from linear principal component analysis of patches of the training images. The training of each of multiple hidden layers CNN includes generating a set of convolutional filters from a selected subset of eigenvectors produced from linear principal component analysis of patches of an affinity matrix constructed using a set of prior convolutional filters from a prior layer of the CNN, where the affinity matrix represents correlations of feature vectors associated with the prior layer. The last layer of the CNN is trained with a regular classifier by error back-propagation using the training images and labels associated with the training images.
1. A method of object recognition using a convolutional neural network (CNN), the method comprising:
a) training a CNN with a set of training images to produce a trained CNN; and
b) classifying an image of an object using the trained CNN;
wherein training the CNN comprises:
for each layer of multiple layers of the CNN, generating a set of convolutional filters associated with the layer from a selected subset of eigenvectors produced from linear principal component analysis of patches of an affinity matrix constructed by convolving the set of training images with a set of prior convolutional filters from a prior layer of the CNN, where the affinity matrix represents correlations of feature vectors associated with the prior layer;
training a regular classifier by error back-propagation using the training images and labels associated with the training images.
2. The method of claim 1 wherein the selected subset of eigenvectors corresponds to eigenvalues larger than a threshold sufficient for the eigenvalues to collectively describe at least 95% of the patches.
3. The method of claim 1 wherein training the CNN further comprises:
for a first layer of the CNN, generating a set of first convolutional filters associated with the first layer from selected first layer eigenvectors produced from linear principal component analysis of patches of the training images.
4. The method of claim 1 wherein training the CNN comprises:
i) performing linear principal component analysis of image patches of the training images to produce first layer eigenvectors and first layer eigenvalues;
ii) selecting a subset of largest first layer eigenvalues and corresponding first layer eigenvectors;
iii) producing from the selected first layer eigenvectors a set of first convolutional filters associated with a first layer of the convolutional neural network;
iv) using the prior convolutional filters produced in the prior layer of the convolutional neural network to produce the feature vectors;
v) constructing the affinity matrix by computing correlations (dot products) between the feature vectors, augmented with the Riemannian distance for co-circularity;
vi) performing linear principal component analysis of patches of the affinity matrix to produce eigenvectors and associated eigenvalues;
vii) selecting the subset of eigenvectors and corresponding eigenvalues by eigenvalue exponentiation and thresholding;
viii) producing from the selected subset of eigenvectors the set of convolutional filters associated with the layer of the convolutional neural network;
ix) repeating steps iv), v), vi), vii), viii) for multiple layers of the convolutional neural network.