Multi-class Learning Model for Image Text Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional OCR technologies are slow and inefficient in identifying text in images, particularly for languages with large alphabets like Chinese and Japanese, making them impractical for real-time anti-spam filtering and content filtering applications.
Innovation Solution
A method using a multi-class learning model trained with a derived list of sensitive characters to extract and identify image character-blocks, converting them into feature vectors, and applying decision trees or neural networks for rapid and accurate prediction of characters, which are then combined into strings for matching against predefined sensitive expressions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional OCR techniques are used to identify text in images with large alphabets, then text extraction is possible, but the process is slow and inefficient
Solution Approach 1:
The patent segments the large alphabet character set into multiple smaller categories (e.g., radicals, strokes, structural components). Instead of training one classifier on all characters, the system divides characters into groups based on their structural features and trains separate classifiers for each group, thereby improving processing speed while maintaining accuracy
Solution Approach 2:
The patent performs preliminary processing on image character-blocks by extracting structural features (radicals, strokes, components) before classification. This preliminary feature extraction and organization allows the classifier to work with pre-processed data, significantly reducing the time required for text identification in real-time applications
2Measurement precision
If OCR is applied to all characters in large alphabets, then complete text recognition is achieved, but computational complexity increases
Solution Approach 1:
The patent segments the character recognition task into multiple stages: structural feature extraction, radical identification, component classification, and final character assembly. This segmentation reduces the complexity of any single classifier while maintaining overall recognition accuracy through a systematic multi-stage approach
Solution Approach 2:
The patent applies different classification strategies to different parts of the character structure. Instead of treating all characters uniformly, the system identifies and classifies specific structural components (radicals, strokes, parts) with specialized classifiers appropriate to each component type, optimizing the model for local characteristics rather than global uniformity
3Object-affected harmful factors
If image-embedded text is used in spam emails, then spam can evade keyword filters, but detection becomes more difficult
Solution Approach 1:
The patent replaces traditional text-based filtering mechanisms with an image processing and recognition system. Instead of searching for keywords in text, the system extracts and recognizes text from images using structural feature analysis and classification, enabling detection of spam content that has been obfuscated by embedding in images
Solution Approach 2:
The patent introduces an intermediary image processing layer between the spam email and the content filter. The system extracts text from images through structural analysis (radicals, strokes, components) before applying content filtering rules, creating an intermediary step that reveals hidden text content while maintaining the ability to detect spam keywords and phrases
Data Source
AI summary
One embodiment relates to a method of identifying sensitive expressions in images for a language with a large alphabet. The method is performed using a computer and includes (i) extracting an image from a message, (ii) extracting image character-blocks (i.e. normalized pixel graphs) from the image, and (iii) predicting characters to which the character-blocks correspond using a multi-class learning model, wherein the multi-class learning model is trained using a derived list of sensitive characters which is a subset of the large alphabet. In addition, (iv) the characters may be combined into string text, and (v) the string text may be searched for matches with a predefined list of sensitive expressions. Another embodiment relates to a method of training a multi-class learning model so that the model predicts characters to which image character-blocks correspond. Other embodiments, aspects and features are also disclosed herein.