Methods and systems for reducing memory footprints associated with classifiers

• US 9,025,865 B2
• Filed: 01/22/2013
• Issued: 05/05/2015
• Est. Priority Date: 09/14/2012
• Status: Active Grant

• Current Assignee: Conduent Business Services, LLC (Conduent, Inc.)
• Original Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Inventors: Kozitsky, Vladimir, Burry, Aaron Michael, Paul, Peter
• Primary Examiner(s): Patel, Jayesh A
• Assistant Examiner(s): TRAN, MAI H
• Application Number:
  US13/746,412
• Publication Number:
  US 20140079315A1
• Time in Patent Office: 833 Days
• Field of Search: None
• US Class Current: 382/159
• CPC Class Codes: A47C 7/14 of adjustable shape; elastically mounted ; adaptable to a user contour or ergonomic seating positions movably-coupled seat and backrests with elastic adjustment means A47C1/03261; elastically-mounted back-rests or backrest-seat units in the base frame A47C7/44
• CPC Class Codes: B60N 2/02 the seat or part thereof being movable, e.g. adjustable adjustable arm-rests B60N2/75; adjustable head-rests B60N2/806
• CPC Class Codes: B60N 2/02253 characterised by the transmission between the electric motor and the seat or seat parts
• CPC Class Codes: B60N 2/1803 with independent front and/or rear adjustment
• CPC Class Codes: B60N 2/1835 pivoting about an axis located in the front
• CPC Class Codes: B60N 2/1864 Gear wheel driven mechanism
• CPC Class Codes: G06F 18/2113 by ranking or filtering the set of features, e.g. using a measure of variance or of feature cross-correlation
• CPC Class Codes: G06F 18/24 Classification techniques
• CPC Class Codes: G06V 10/771 Feature selection, e.g. selecting representative features from a multi-dimensional feature space
• CPC Class Codes: G06V 20/625 License plates
• CPC Class Codes: G06V 20/63 Scene text, e.g. street names
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/18133 regional/local feature not essentially salient, e.g. local binary pattern

Abstract

Methods and systems for reducing the required footprint of SNoW-based classifiers via optimization of classifier features. A compression technique involves two training cycles. The first cycle proceeds normally and the classifier weights from this cycle are used to rank the Successive Mean Quantization Transform (SMQT) features using several criteria. The top N (out of 512 features) are then chosen and the training cycle is repeated using only the top N features. It has been found that OCR accuracy is maintained using only 60 out of 512 features leading to an 88% reduction in RAM utilization at runtime. This coupled with a packing of the weights from doubles to single byte integers added a further 8× reduction in RAM footprint or a reduction of 68× over the baseline SNoW method.

Claims

1. A method for reducing the required footprints associated with classifiers, said method comprising:

employing classifier weights associated with SNoW classifiers from a SNoW classifier training cycle to rank utilizing particular criteria features comprising SMQT features;

selecting at least one top feature among said SMQT features;

repeating said SNoW classifier training cycle using only said at least one top feature in order to thereafter determine if additional features among said SMQT features should be included or excluded to reduce said footprints associated with said SNoW classifiers; and

wherein a size of said SNoW classifiers is compressed in order to identify important features among said features and wherein for each pixel location said classifier weights are assigned to each of said features as part of said SNoW classifier training cycle and wherein during a recognition phase only feature among said features is active at said each pixel location based on an SMQT bit pattern in a location neighborhood and wherein a sum of active classifier weights among said classifier weights is employed to determine a resultant class label for a test image.

2. The method of claim 1 wherein said classifier weights are contained in a 3D classifier matrix and wherein said classifier training cycle manipulates said classifier weights of said 3D matrix so that all characters in training are correctly categorized.

3. The method of claim 1 further comprising:

analyzing a 3D classifier matrix associated with said classifiers to determine a relative importance of said features across all spatial locations; and

ranking said features by employing said particular criteria comprising at least one of all of the following;

a count of non-zero weights with respect to said feature, said feature comprising at least a particular feature plane;

a mean/median value of weights associated with a feature plane of said feature; and

a highest weight feature plane associated with said feature with respect to at least one pixel location with a highest value of non-zero weight.

4. The method of claim 1 further comprising ranking said features by employing said particular criteria, wherein said ranking of said features is custom to each character classifier in an OCR engine or said ranking is carried across all of said classifiers to determine a globally optimal set of features.

5. The method of claim 4 wherein said particular criteria comprises a count of non-zero weights with respect to said feature, said feature comprising at least a particular feature plane.

6. The method of claim 5 wherein said particular criteria further comprises a mean/median value of weights associated with a feature plane of said feature.

7. The method of claim 5 wherein said particular criteria further comprises a highest weight feature plane associated with said feature with respect to at least one pixel location with a highest value of non-zero weight.

8. A system for reducing the required footprints associated with classifiers, said system comprising:

a processor;

a data bus coupled to said processor;

a non-transitory computer-usable medium embodying computer program code, said computer-usable medium being coupled to said data bus, said computer program code comprising instructions executable by said processor and configured for;

employing classifier weights associated with SNoW classifiers from a SNoW training cycle to rank utilizing particular criteria features comprising SMQT features;

selecting at least one top feature among said SMQT features;

repeating said SNoW classifier training cycle using only said at least one top feature among said SMQT features in order to thereafter determine if additional features among said SMQT features should be included or excluded to reduce said footprints associated with said SNoW classifiers; and

wherein a size of said SNoW classifiers is compressed in order to identify important features among said features and wherein for each pixel location said classifier weights are assigned to each of said features as part of said SNoW classifier training cycle and wherein during a recognition phase only feature among said features is active at said each pixel location based on an SMQT bit pattern in a location neighborhood and wherein a sum of active classifier weights among said classifier weights is employed to determine a resultant class label for a test image.

9. The system of claim 8 wherein said classifier weights are contained in a 3D classifier matrix and wherein said classifier training cycle manipulates said classifier weights of said 3D matrix so that all characters in training are correctly categorized.

10. The system of claim 8 wherein said instructions are further configured for:

analyzing a 3D classifier matrix associated with said classifiers to determine a relative importance of said features across all spatial locations; and

ranking said features by employing said particular criteria comprising at least one of all the following;

a count of non-zero weights with respect to said feature, said feature comprising at least a particular feature plane;

a mean/median value of weights associated with a feature plane of said features; and

a highest weight feature plane associated with said feature with respect to at least one pixel location with a highest value of non-zero weight.

11. The system of claim 9 wherein for each pixel location said classifier weights are assigned to each of said features as part of said SNoW classifier training cycle.

12. The system of claim 9 wherein a size of said SNoW classifiers is compressed in order to identify important features among said features.

13. The system of claim 11 wherein during a recognition phase only feature among said features is active at said each pixel location based on an SMQT bit pattern in a location neighborhood.

14. The system of claim 11 wherein a sum of active classifier weights among said classifier weights is employed to determine a resultant class label for a test image.

15. A non-transitory processor-readable medium storing code representing instructions to cause a process for reducing the required footprints associated with classifiers, said code comprising code to:

employ classifier weights associated with SNoW classifiers from a SNoW classifier training cycle to rank utilizing particular features comprising SMQT features;

select at least one top feature among said features;

repeat said SNoW classifier training cycle using only said at least one top feature in order to thereafter determine if additional features among said features should be included or excluded to reduce said footprints associated with said classifiers; and

wherein a size of said SNoW classifiers is compressed in order to identify important features among said features and wherein for each pixel location said classifier weights are assigned to each of said features as part of said SNoW classifier training cycle and wherein during a recognition phase only feature among said features is active at said each pixel location based on an SMQT bit pattern in a location neighborhood and wherein a sum of active classifier weights among said classifier weights is employed to determine a resultant class label for a test image.

16. The non-transitory processor-readable medium of claim 15 wherein said classifier weights are contained in a 3D classifier matrix and wherein said classifier training cycle manipulates said classifier weights of said 3D matrix so that all characters in training are correctly categorized.

17. The non-transitory processor-readable medium of claim 15 wherein said code further comprises code to:

analyze a 3D classifier matrix associated with said SNoW classifiers to determine a relative importance of said features across spatial locations; and

rank said features by employing said particular criteria comprising at least one of all of the following;

a count of non-zero weights with respect to said feature, said feature comprising at least a particular feature plane;

a mean/median value of weights associated with a feature plane of said feature; and

a highest weight feature plane associated with said feature with respect to at least one pixel location with a highest value of non-zero weight.