System and method for scene text recognition

Abstract

Apparatus and method for performing accurate text recognition of non-simplistic images (e.g., images with clutter backgrounds, lighting variations, font variations, non-standard perspectives, and the like) may employ a machine-learning approach to identify a discriminative feature set selected from among features computed for a plurality of irregularly positioned, sized, and/or shaped (e.g., randomly selected) image sub-regions.

1 . (canceled) 
     
     
         2 . A system comprising:
 one or more computer processors; and   memory storing instructions that, when executed by the one or more computer processors, cause the one or more computer processors to perform operations comprising:   computing, from each of a plurality of training images having specified image dimensions, features for a plurality of irregularly sized and positioned sub-regions defined for an image region having the specified image dimensions;   determining feature weights associated with the computed features,   ranking the features based on the feature weights; and   selecting, among the features, a number of top-ranked features for inclusion in a discriminative feature space.   
     
     
         3 . The system of  claim 1 , wherein the operations comprise use of machine learning to determine the feature weights. 
     
     
         4 . The system of  claim 1 , wherein the operations further comprise computing pixel-wise low-level features from the plurality of training images, the features for the plurality of sub-regions being mid-level features computed by aggregating the low-level features over the pixels within the respective sub-regions. 
     
     
         5 . The system of  claim 1 , wherein the plurality of training images are associated with a plurality of character classes, and wherein the operations comprise determining the feature weights, ranking the features, and selecting the top-ranked features separately for each of the plurality of character classes, the operations further comprising combining the selected top-ranked features across character classes into a combined feature space. 
     
     
         6 . The system of  claim 5 , wherein the operations further comprise updating the feature weights associated with the top-ranked features combined into the combined feature space. 
     
     
         7 . The system of  claim 5 , wherein the operations further comprise classifying a scene character based on an image thereof by computing features within the combined feature space from the image of the scene character, computing weighted averages of the features for the plurality of character classes, and comparing the weighted averages to identify a closest match for the scene character among the plurality of character classes. 
     
     
         8 . The system of  claim 1 , wherein the operations further comprise defining the plurality of irregularly sized and positioned sub-regions by randomly selecting positions, widths, and heights from respective uniform distributions. 
     
     
         9 . The system of  claim 1 , wherein the features comprise features computed for a plurality of feature channels. 
     
     
         10 . The system of  claim 8 , wherein the feature channels comprises at least one of a color channel, a gradient histogram channel, and a gradient-magnitude channel. 
     
     
         11 . A non-transitory machine-readable medium storing a plurality of processor-executable instructions which, when executed by one or more processors of the machine, cause the one or more processors to perform operations comprising:
 computing, from each of a plurality of training images having specified image dimensions, features for a plurality of irregularly sized and positioned sub-regions defined for an image region having the specified image dimensions;   determining feature weights associated with the computed features;   ranking the features based on the feature weights; and   selecting, among the features, a number of top-ranked features for inclusion in a discriminative feature space.   
     
     
         12 . The machine-readable medium of  claim 11 , wherein the operations comprise use of machine learning to determine the feature weights. 
     
     
         13 . The machine-readable medium of  claim 11 , wherein the operations further comprise computing pixel-wise low-level features from the plurality of training images, the features for the plurality of sub-regions being mid-level features computed by aggregating the low-level features over the pixels within the respective sub-regions. 
     
     
         14 . The machine-readable medium of  claim 11 , wherein the plurality of training images are associated with a plurality of character classes, and wherein the operations comprise determining the feature weights, ranking the features, and selecting the top-ranked features separately for each of the plurality of character classes, the operations further comprising combining the selected top-ranked features across character classes into a combined feature space. 
     
     
         15 . The machine-readable medium of  claim 14 , wherein the operations further comprise updating the feature weights associated with the top-ranked features combined into the combined feature space. 
     
     
         16 . The machine-readable medium of  claim 14 , wherein the operations further comprise classifying a scene character based on an image thereof by computing features within the combined feature space from the image of the scene character, computing weighted averages of the features for the plurality of character classes, and comparing the weighted averages to identify a closest match for the scene character among the plurality of character classes. 
     
     
         17 . The machine-readable medium of  claim 11 , wherein the operations further comprise defining the plurality of irregularly sized and positioned sub-regions by randomly selecting positions, widths, and heights from respective uniform distributions. 
     
     
         18 . A method comprising:
 computing, from each of a plurality of training images having specified image dimensions, features for a plurality of irregularly sized and positioned sub-regions defined for an image region having the specified image dimensions;   determining feature weights associated with the computed features;   ranking the features based on the feature weights; and   selecting, among the features, a number of top-ranked features for inclusion in a discriminative feature space.   
     
     
         19 . The method of  claim 18 , wherein the plurality of training images are associated with a plurality of character classes, and wherein the feature weights are determined, the features are ranked, and top-ranked features are selected separately for each of the plurality of character classes, the method further comprising combining the selected top-ranked features across character classes into a combined feature space. 
     
     
         20 . The method of  claim 19 , further comprising classifying a scene character based on an image thereof by computing features within the combined feature space from the image of the scene character, computing weighted averages of the features for the plurality of character classes, and comparing the weighted averages to identify a closest match for the scene character among the plurality of character classes. 
     
     
         21 . The method of  claim 18 , further comprising defining the plurality of irregularly sized and positioned sub-regions by randomly selecting positions, widths, and heights from respective uniform distributions.