Systems and methods for detecting text

• US 7,570,816 B2
• Filed: 03/31/2005
• Issued: 08/04/2009
• Est. Priority Date: 03/31/2005
• Status: Expired due to Fees

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Viola, Paul A, Simard, Patrice Y, Bargeron, David M
• Primary Examiner(s): Mehta; Bhavesh M
• Assistant Examiner(s): Akhavannik; Hadi
• Application Number:
  US11/095,393
• Publication Number:
  US 20060222239A1
• Time in Patent Office: 1,587 Days
• Field of Search: 382/224, 382/176
• US Class Current: 382/224
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06V 30/19147 Obtaining sets of training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06V 30/40 Document-oriented image-based pattern recognition

Abstract

The subject invention relates to facilitating text detection. The invention employs a boosted classifier and a transductive classifier to provide accurate and efficient text detection systems and/or methods. The boosted classifier is trained through features generated from a set of training connected components and labels. The boosted classifier utilizes the features to classify the training connected components, wherein inferred labels are conveyed to a transductive classifier, which generates additional properties. The initial set of features and the properties are utilized to train the transductive classifier. Upon training, the system and/or methods can be utilized to detect text in data under text detection, wherein unlabeled data is received, and connected components are extracted therefrom and utilized to generate corresponding feature vectors, which are employed to classify the connected components using the initial boosted classifier. Inferred labels are utilized to generate properties, which are utilized along with the initial feature vectors to classify each connected component using the transductive classifier.

Claims

1. A system, embodied on a computer-readable storage medium, that when executed, facilitates detecting text in data, comprising:

an input component that receives data;

a connected components identifier that generates a set of connected components from the data, the connected components are utilized to generate a set of features for training and testing; and

a classification component that automatically detects text in the data via a transductive classifier employed in connection with a trained boosted classifier, the trained boosted classifier infers labels for the training connected components, the inferred labels are subjected to a clustering process by which the set of training features feature sets are expanded to define training properties, the transductive classifier is trained based in part upon the training properties.

2. The system of claim 1, the connected components identifier utilizes a union-find algorithm to identify the training connected components.

3. The system of claim 1, further comprising a feature generator that generates a set of training features that are employed to train the boosted classifier and the transductive classifier.

4. The system of claim 3, the feature generator utilizes one of a spatial relation technique or an activity density technique to generate the set of features.

5. The system of claim 1, the data is one of a document, a spreadsheet, an image, a page, a web page, a video frame, or a drawing.

6. An optical character recognition (OCR) system that employs the system of claim 1 in connection with extracting text from an image.

7. A method for detecting text, comprising:

employing a processor to execute computer executable code stored on a storage medium to perform the following acts;

identifying one or more connected components associated with unlabeled data under text detection;

utilizing the connected components to extract a feature vector for each connected component;

utilizing a boosted classifier to classify each connected component represented by its respective feature vector;

employing inferred labels to bin the connected components across a plurality of bins;

computing properties for each bin; and

utilizing a transductive classifier to classify the connected components as a function of the feature vectors and corresponding computed bin properties.

8. The method of claim 7, further comprising training the boosted classifier by:

receiving labeled training data that includes connected components, text labels and nontext labels;

extracting a feature vector for each connected component in the training data; and

training the boosted classifier with the labels and feature vectors.

9. The method of claim 8, further comprising training the transductive classifier by:

employing the boosted classifier to classify each training connected component, represented by a respective feature vector;

utilizing labels inferred from the training connected components to bin the training connected components across a plurality of bins;

computing properties for each bin; and

utilizing the labels, feature vectors and properties to train the transductive classifier.

10. The method of claim 9, further comprising utilizing one of a spatial relation technique or an activity density technique to generate the feature vectors for the unlabeled data and the labeled training data.

11. The method of claim 9, the bins utilized to delineate the connected components of the unlabeled data and the training data are associated with one or more histograms that correspond to at least one of the following:

a number of pixels in a connected component;

a height of a connected component'"'"'s bounding box;

a width of a connected component'"'"'s bounding box;

or a number of pixels in a connected component divided by a connected component'"'"'s bounding box area.

12. The method of claim 8, the transductive classifier is a second boosted classifier.

13. The method of claim 8, further comprising employing a decision tree with fixed depth in connection with the boosted classifier and the transductive classifier to allow for complex conditional processing of raw features without over-fitting the data.

18. The method of claim 9, further comprising providing information to an optical character recognition (OCR) system utilized to at least one of identify or extract the detected text from the unlabeled data.

14. A system, embodied on a computer-readable storage medium, that when executed, trains a text detector, comprising:

means for identifying a set of connected components from at least one of unlabeled data or labeled data;

means for extracting a feature vector for each connected component identified;

means for training a boosted classifier with the connected components and corresponding text and non-text labels for text detection;

means for training a transductive classifier with as a function of an expanded feature set, the expanded feature set generated from at least one of labels, feature vectors and computed bin properties; and

means for employing the trained boosted classifier in connection with the trained transductive classifier to detect text within the unlabeled data.

15. A computer-readable storage medium having computer-executable instructions stored thereon to perform a method comprising:

receiving labeled training data that includes connected components, text labels and nontext labels;

identifying one or more connected components associated with the labeled training data;

utilizing a spatial relation technique to generate a feature vector for each of the one or more connected components identified in the labeled data;

training one or more boosted classifiers with the labels and the feature vectors;

employing the one or more trained boosted classifiers to infer labels for the training connected components;

generating one or more histograms that define a plurality of equally sized bins as a function of a percentage of a maximum range of feature values;

utilizing the labels inferred from the training connected components to bin the training connected components across the plurality of bins;

computing properties and additional features for each bin;

generating expanded feature vectors with the properties and additional features;

utilizing the original training data labels and expanded feature vectors to train a transductive classifier;

receiving unlabeled data; and

employing the trained boosted classifier and the trained transductive classifier to detect text within the unlabeled data.

16. The computer-readable storage medium of claim 15, further comprising:

providing information to an optical character recognition (OCR) system utilized to at least one of identify or extract the detected text from the unlabeled data.

17. The computer-readable storage medium of claim 15, the unlabeled data is one of a document, a spreadsheet, an image, a page, a web page, a video frame or a drawing.