Layout-Aware OCR Feature Vector for Noisy Text Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional optical character recognition (OCR) systems in recognition systems produce noisy outputs due to imperfections in images, leading to misrecognized characters and words, and lack robustness in handling such noise, which affects the accuracy of recognition processes.
Innovation Solution
The implementation of a layout-aware text feature that uses 3-letter-grams and incorporates 2D layout information from OCR outputs, along with normalization and matching schemes, to enhance robustness against spelling errors and misrecognitions, and employs a retrieval approach instead of classification to handle large datasets effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional OCR is used for text recognition, then the recognition process can be implemented, but the output becomes noisy with misrecognized characters and words
Solution Approach 1:
The patent segments the text recognition problem into multiple components: extracting text blocks with positional information, generating n-grams from recognized text, creating frequency vectors, and computing similarity scores. This segmentation allows each component to be optimized independently and combines them to overcome individual weaknesses, particularly addressing OCR noise through the multi-stage processing pipeline.
Solution Approach 2:
The patent introduces intermediate representations as mediators between OCR output and final recognition results. These include text block extractions with positional properties, n-gram sequences, frequency vectors, and similarity scores. These intermediaries allow the system to process and refine OCR output through multiple transformation stages, converting noisy character-level OCR into more robust document-level similarity measurements.
2Adaptability or versatility
If classification approach is used for recognition, then model can learn from data, but it requires frequent updates and does not scale well to billions of documents
Solution Approach 1:
The patent inverts the traditional classification approach by using similarity search instead. Rather than training a model to classify documents into categories, the system converts documents into frequency vectors and retrieves similar documents through vector similarity computation. This inversion eliminates the need for frequent model retraining while maintaining adaptability, as new documents are simply added to the vector index without requiring model updates.
Solution Approach 2:
The patent changes the fundamental parameter representation from class labels to continuous frequency vectors. By representing documents as vectors of n-gram frequencies rather than discrete categories, the system enables efficient similarity search and scaling to billions of documents. This parameter transformation allows the use of approximate nearest neighbor search algorithms that scale linearly or near-linearly with document count, unlike quadratic classification approaches.
3Device complexity
If layout information is ignored in OCR processing, then processing is simpler, but recognition accuracy deteriorates due to noisy output
Solution Approach 1:
The patent adds the spatial dimension to text processing by incorporating positional properties (x, y coordinates, width, height) of text blocks alongside the textual content. This dimensional expansion allows the system to distinguish between text from different locations in the document, preserving structural information that helps disambiguate OCR errors. The positional dimension is integrated into the feature vector, creating a more robust representation that combines textual and spatial characteristics.
Data Source
AI summary
Described herein is a mechanism for visual recognition of items or visual search using Optical Character Recognition (OCR) of text in images. Recognized OCR blocks in an image comprise position information and recognized text. The embodiments utilize a location-aware feature vector created using the position and recognized information in each recognized block. The location-aware features of the feature vector utilize position information associated with the block to calculate a weight for the block. The recognized text is used to construct a tri-character gram frequency, inverse document frequency (TGF-IDF) metric using tri-character grams extracted from the recognized text. Features in location-aware feature vector for the block are computed by multiplying the weight and the corresponding TGF-IDF metric. The location-aware feature vector for the image is the sum of the location-aware feature vectors for the individual blocks.