Image quality assessment using similar scenes as reference
A system for image quality assessment of non-aligned images includes a first deep path portion of a convolutional neural network having a set of parameters and a second deep path portion of the convolutional neural network sharing a set of parameters with the first deep path convolutional neural network. Weights are shared between the first and second deep path convolutional neural networks to support extraction of a same set of features in each neural network pathway. Non-aligned reference and distorted images are respectively provided to the first and second deep paths of the convolutional neural network for processing. A concatenation layer is connected to both the first and second deep paths convolutional neural network, and a fully connected layer is connected to the concatenation layer to receive input from both the first and second deep paths of the convolutional neural network, generating an image quality assessment as a linear regressor and outputting an image quality score.
1. A method to predict an image quality score associated with a plurality of images, the method comprising:
receiving, by a first deep path convolutional neural network, a set of reference images;
receiving, by a second deep path convolutional neural network, a set of distorted images, wherein each image in the set of distorted images is non-aligned with each image in the set of reference images;
extracting, by the first deep path convolutional neural network, a first set of one or more local features from the set of reference images;
extracting, by the second deep path convolutional neural network, a second set of one or more local features from the set of distorted images;
concatenating the first set of local features and the second set of local features;
generating an image quality assessment; and
predicting an image quality score.
2. The method of claim 1 , wherein the extracting by the first deep path convolutional neural network is performed using a first convolution operation, and wherein the extracting by the second deep path convolutional neural network is performed using a second convolution operation.
3. The method of claim 2 , wherein the first convolution operation is performed by a first convolution layer, and wherein the second convolution operation is performed by a second convolution layer.
4. The method of claim 3 , wherein each of the first convolution layer and the second convolution layer is trained to extract one or more local features from an image.
5. The method of claim 1 , further comprising:
comparing the first set and the second set; and
compensating for an offset between the set of reference images and the set of distorted images.
6. The method of claim 5 , wherein the comparing is performed using a weight sharing strategy.
7. The method of claim 5 , wherein the compensating is performed using one or more pool layers.
8. The method of claim 7 , further comprising performing a max-pooling operation.
9. The method of claim 1 , wherein the image quality assessment is generated using a linear regressor.
10. The method of claim 1 , wherein the image quality score is predicted by minimizing a Euclidian loss function.
11. The method of claim 1 , further comprising training each of the first deep path convolutional neural network and the second deep path convolutional neural network.
12. The method of claim 11 , further comprising accelerating the training using a rectified linear unit as an activation function.
13. The method of claim 11 , wherein the training is performed using a stochastic gradient descent with standard back-propagation.
14. An apparatus comprising:
at least one camera system providing distorted images; and
a computer system including a first deep path convolutional neural network and a second deep path convolutional neural network, the first deep path convolutional neural network accessing reference images while using a set of parameters, the second deep path convolutional neural network receiving the distorted images from the at least one camera system, wherein each image in the set of distorted images is non-aligned with each image in the set of reference images, the first deep path convolutional neural network extracting a first set of one or more local features from the set of reference images, the second deep path convolutional neural network extracting a second set of one or more local features from the set of distorted images, the computer system concatenating the first set of local features and the second set of local features, the computer system generating an image quality assessment, and the computer system predicting an image quality score.
15. The apparatus of claim 14 , wherein the extracting by the first deep path convolutional neural network is performed using a first convolution operation, and wherein the extracting by the second deep path convolutional neural network is performed using a second convolution operation.
16. The apparatus of claim 15 , wherein the first convolution operation is performed by a first convolution layer, and wherein the second convolution operation is performed by a second convolution layer.
17. The apparatus of claim 16 , wherein each of the first convolution layer and the second convolution layer is trained to extract one or more local features from an image.
18. The apparatus of claim 1 , further comprising the computer system:
comparing the first set and the second set; and
compensating for an offset between the set of reference images and the set of distorted images.
19. The apparatus of claim 18 , wherein the comparing is performed using a weight sharing strategy.
20. The apparatus of claim 18 , wherein the compensating is performed using one or more pool layers.
21. The apparatus of claim 20 , further comprising performing a max-pooling operation.
22. The apparatus of claim 14 , wherein the image quality assessment is generated using a linear regressor.
23. The apparatus of claim 14 , wherein the image quality score is predicted by minimizing a Euclidian loss function.
24. The apparatus of claim 14 , further comprising training each of the first deep path convolutional neural network and the second deep path convolutional neural network.
25. The apparatus of claim 24 , further comprising accelerating the training using a rectified linear unit as an activation function.
26. The apparatus of claim 24 , wherein the training is performed using a stochastic gradient descent with standard back-propagation.