Method for classification of unique/rare cases by reinforcement learning in neural networks

1. A method of image processing, comprising:

capturing image data as original data;

supplying unlabeled images as unlabeled data on a higher scale than the original data;

reinforcing deep neural network learning capacity to classify rare cases amongst the captured image data, by:

training a first deep neural network by:

supplying the original data containing generic and rare cases and specified labels; and

outputting at least one class-specific feature space and at least one class-specific decision space used to classify generic cases of the original data into the specified labels;

localizing discriminative class-specific features within the original data processed through the at least one class-specific feature spaces of the first deep neural networks and mapping the discriminative class-specific features on the original data as spatial-probabilistic labels;

training a second-deep neural network by:

supplying the original data and the spatial-probabilistic labels; and

outputting at least one universal feature space and at least one universal decision space used to the classify rare cases of the original data into the spatial-probabilistic labels;

training a combined deep neural network by:

supplying the original data, the specified labels and the spatial-probabilistic labels; and

outputting primary combined class-specific and universal feature and decision spaces used to classify both the generic and rare cases of the original data into primary combined specified and spatial-probabilistic labels;

further training of the combined deep neural network by:

supplying unlabeled data;

classifying the unlabeled data into the primary combined specified and spatial-probabilistic labels;

mapping the primary combined specified and spatial-probabilistic labels into secondary combined specified and spatial-probabilistic misclassification labels;

training a second combined deep neural network by:

supplying the unlabeled data, the secondary misclassification labels and a penalty matrix; and

outputting secondary combined decision and feature spaces used to classify both generic and rare cases of the unlabeled data into the secondary misclassification labels (Label-C) according to the penalty matrix classifying the generic and rare cases based on the captured images.

2. The method of image processing of claim 1 , wherein the first, second and combined deep neural networks are convolutional neural networks.

3. The method of image processing of claim 1 , wherein the localizing step of discriminative class-specific features includes:

randomly extracting patches of the original data processed through the at least one class-specific feature spaces;

ranking randomly extracted patches by their statistical popularity among all patches of a same specified label;

bagging visual words features starting from the top-ranked patches;

random forest tree classifying of bag of visual words features; and

extracting variable importance measures using the random forest tree classification to localize discriminative class-specific features and estimate their relative importance.

4. The method of image processing of claim 1 , wherein the training step of the second-deep neural network further includes:

resizing spatial-probabilistic labels;

creating batches of the original data with corresponding resized spatial-probabilistic labels to learn universal feature and decision spaces via back-propagation of feedback generated by a universal loss function;

estimating localization and probability of the discriminative class-specific features independently from the specified labels; and

classifying rare cases of the original data into the resized spatial-probabilistic labels.

5. The method of image processing of claim 1 , wherein the training step of the combined deep neural network further includes:

combining the at least one universal feature space and the at least one class-specific feature space into at least one integrated feature space forming the primary combined class-specific and universal feature space with different learning rates;

combining the at least one universal decision space and the at least one class-specific decision space into at least one integrated decision space forming the primary combined class-specific and universal decision space; and

using a combined loss function of the universal loss function of the second deep neural network and a class-specific loss function of the first deep neural network to provide feedback for original data belonging to any label out of the primary combined specified and spatial-probabilistic labels.

6. The method of image processing of claim 5 , wherein the learning rate of the universal feature space is smaller than the learning rate of the class-specific feature space.

7. The method of image processing of claim 1 , wherein the training step of the second combined deep neural network further includes:

starting from the trained combined deep neural network;

extending feature and decision spaces by adding hidden layers with a purpose to increase network's capacity to host newly learned features and decisions; and

modifying a loss function of the second combined deep neural network or batches of unlabeled data by using more data samples for the rare cases, with higher penalty on misclassification based on the penalty matrix.

8. The method of image processing of claim 7 , further comprising:

training a third combined deep neural network by:

supplying the original data containing the generic and rare cases, the specified labels and the penalty matrix; and

outputting primary combined decision and feature spaces used to classify both the generic and rare cases of the original data into the specified labels according to the penalty matrix.

9. The method of image processing of claim 8 , wherein the training step of the third combined deep neural network further includes:

starting from the trained second combined deep neural network on the unlabeled data and the secondary misclassification labels;

extending classify the rare cases by adding hidden layers; and

modifying a loss function of the third combined deep neural network or batches of unlabeled data by using more data samples for the rare cases, with higher penalty on misclassification based on the penalty matrix.

10. The method of image processing according to claim 8 , wherein:

the mapping step of primary combined labels into secondary combined misclassification labels together with the training steps of both the second and third combined deep neural networks form a small reinforcement loop; and

the steps from the training step of the first deep neural network to the training step of the third combined deep neural network form a big reinforcement loop.

11. A vehicle, comprising:

a path capturing unit configured to capture and convert portions of a followed path seen at least from a driver's point of view into a series of digital files, when the vehicle is driven,

a processor configured to process the series of digital files, and to classify generic and rare cases based on the series of digital files according to the image processing method of claim 10 ;

a display configured to display an information related to the classified generic and rare cases;

an autonomous driving unit configured to control the vehicle; and

a decision unit configured to activate at least one of the display unit and the autonomous driving unit depending on the classified rare cases.

12. The method of image processing of claim 2 , wherein the localizing step of discriminative class-specific features includes:

randomly extracting patches of the original data processed through the at least one class-specific feature spaces;

ranking randomly extracted patches by their statistical popularity among all patches of a same specified label;

bagging visual words features starting from the top-ranked patches;

random forest tree classifying of bag of visual words features; and

extracting variable importance measures using the random forest tree classification to localize discriminative class-specific features and estimate their relative importance.

13. The method of image processing of claim 2 , wherein the training step of the second-deep neural network further includes:

resizing spatial-probabilistic labels;

creating batches of original data with corresponding resized spatial-probabilistic labels to learn universal feature and decision spaces via back-propagation of feedback generated by a universal loss function;

estimating localization and probability of the discriminative class-specific features independently from specified labels; and

classifying rare cases of the original data into resized spatial-probabilistic labels.

14. The method of image processing of claim 3 , wherein the training step of the second-deep neural network further includes:

resizing spatial-probabilistic labels;

creating batches of original data with corresponding resized spatial-probabilistic labels to learn universal feature and decision spaces via back-propagation of feedback generated by a universal loss function;

estimating localization and probability of the discriminative class-specific features independently from specified labels; and

classifying rare cases of the original data into resized spatial-probabilistic labels.

15. The method of image processing of claim 2 , wherein the training step of the combined deep neural network further includes:

combining at least one universal feature space and at least one class-specific feature space into at least one integrated feature space forming the primary combined class-specific and universal feature space with different learning rates;

combining at least one universal decision space and at least one class-specific decision space into at least one integrated decision space forming the primary combined class-specific and universal decision space; and

using a combined loss function of the universal loss function of the second-deep neural network and a class-specific loss function of the first deep neural network to provide feedback for original data belonging to any label out of the primary combined specified and spatial-probabilistic labels.

16. The method of image processing of claim 3 , wherein the training step of the combined deep neural network further includes:

combining at least one universal feature space and at least one class-specific feature space into at least one integrated feature space forming the primary combined class-specific and universal feature space with different learning rates;

combining at least one universal decision space and at least one class-specific decision space into at least one integrated decision space forming the primary combined class-specific and universal decision space; and

using a combined loss function of the universal loss function of the second-deep neural network and a class-specific loss function of the first deep neural network to provide feedback for original data belonging to any label out of the primary combined specified and spatial-probabilistic labels.

17. The method of image processing of claim 4 , wherein the training step of the combined deep neural network further includes:

combining at least one universal feature space and at least one class-specific feature space into at least one integrated feature space forming the primary combined class-specific and universal feature space with different learning rates;

combining at least one universal decision space and at least one class-specific decision space into at least one integrated decision space forming the primary combined class-specific and universal decision space; and

using a combined loss function of the universal loss function of the second-deep neural network and a class-specific loss function of the first deep neural network to provide feedback for original data belonging to any label out of the primary combined specified and spatial-probabilistic labels.

18. The method of image processing of claim 2 , wherein the training step of the second combined deep neural network further includes:

starting from trained combined deep neural network;

extending feature and decision spaces by adding hidden layers with a purpose to increase network's capacity to host newly learned features and decisions; and

modifying a loss function of the second combined deep neural network or batches of unlabeled data by using more data samples for the rare cases, with higher penalty on misclassification based on the penalty matrix.

19. The method of image processing of claim 3 , wherein the training step of the second combined deep neural network further includes:

starting from trained combined deep neural network;

extending feature and decision spaces by adding hidden layers with a purpose to increase network's capacity to host newly learned features and decisions; and

modifying a loss function of the second combined deep neural network or batches of unlabeled data by using more data samples for the rare cases, with higher penalty on misclassification based on the penalty matrix.

20. The method of image processing of claim 4 , wherein the training step of the second combined deep neural network further includes:

starting from trained combined deep neural network;

extending feature and decision spaces by adding hidden layers with a purpose to increase network's capacity to host newly learned features and decisions; and

modifying a loss function of the second combined deep neural network or batches of unlabeled data by using more data samples for the rare cases, with higher penalty on misclassification based on the penalty matrix.