Optical Character Recognition Confidence Functions for Grapheme Validation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks for optical character recognition (OCR) face challenges in distinguishing valid graphemes from invalid or indicia-like images due to insufficiently specific automatically defined feature sets and the scarcity of negative samples, leading to errors in image segmentation and recognition.
Innovation Solution
The implementation of specialized confidence functions that compute distances between feature vectors and class centers in the image feature space, transforming these distances into confidence values to improve the recognition of valid graphemes and filter out invalid images, using a combination of Cross Entropy Loss, Center Loss, and Close-to-Center Penalty Loss functions during neural network training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If traditional neural networks use automatically defined feature sets for OCR, then the system can process images without manual feature engineering, but the recognition accuracy deteriorates due to insufficiently specific features and inability to distinguish valid graphemes from invalid images
Solution Approach 1:
The patent transforms the feature representation by computing distances between feature vectors and class centers in the image feature space. This parameter transformation converts raw feature vectors into confidence values that directly indicate recognition certainty, resolving the contradiction by changing how features are utilized rather than how they are defined
Solution Approach 2:
The patent introduces class centers as intermediary reference points in the feature space. These class centers serve as mediators between the automatically defined feature vectors and the final recognition decision, enabling accurate distinction between valid and invalid graphemes while maintaining automatic feature definition
2Productivity
If the neural network is trained with traditional loss functions, then the training process is simple and fast, but the system fails to filter out invalid images and produces segmentation errors
Solution Approach 1:
The patent performs preliminary computation of class centers during the training process using training data, before using these centers for confidence computation on validation and test data. This preliminary action enables the system to establish reference points for valid graphemes in advance, improving reliability without significantly impacting training speed
Solution Approach 2:
The patent implements a feedback mechanism where confidence values computed from distances to class centers are used to evaluate and filter recognition results. This feedback loop allows the system to identify and reject low-confidence predictions, thereby filtering out invalid images while maintaining efficient training
Data Source
AI summary
Systems and methods for optical character recognition using specialized confidence functions. An example method comprises: receiving a grapheme image; computing a feature vector representing the grapheme image in a space of image features; and computing a confidence vector associated with the grapheme image, wherein each element of the confidence vector reflects a distance, in the space of image features, between the feature vector and a center of a class of a set of classes.