Multi-Type Document Detection With Shared Neural Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing document processing technologies are less optimal for comprehensive detection and classification of multiple object types in documents, requiring multiple algorithmic processes and inefficient use of computational resources.
Innovation Solution
A system and method utilizing a backbone neural network and classification heads to generate pixel-level and object-level maps for concurrent detection of various object types in documents, including preprocessing techniques like rescaling, cropping, and padding, with a machine learning model that trains classification heads end-to-end.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple separate algorithmic processes are used for detection and classification of different object types, then detection coverage is improved, but device complexity and computational resource usage increase
Solution Approach 1:
The patent combines multiple separate algorithmic processes into a single machine learning model that performs detection and classification of multiple object types (text, graphics, tables, forms) simultaneously. The model uses a unified architecture with shared feature extraction layers and multiple classification heads, eliminating the need for separate processing pipelines for each object type while maintaining comprehensive detection coverage.
Solution Approach 2:
The machine learning model is designed as a universal system that can detect and classify multiple types of document objects within a single framework. The model accepts various object types as input and produces unified output, making it adaptable to different document contents without requiring separate specialized algorithms for each object category.
2Adaptability or versatility
If multiple separate algorithmic processes are used for detection and classification, then comprehensive detection is improved, but processing time and computational resources increase
Solution Approach 1:
By merging multiple detection and classification algorithms into a single machine learning model, the system processes all object types in one pass through the network. The unified architecture shares computational resources across different object types, eliminating redundant processing steps and significantly reducing overall processing time compared to running separate algorithms sequentially.
Solution Approach 2:
The model performs preliminary feature extraction that serves all subsequent classification tasks. By extracting general document features once at the beginning and then using them for multiple classification heads, the system avoids redundant processing and improves processing speed while maintaining comprehensive detection capabilities.
3Adaptability or versatility
If traditional OCR and object recognition methods are used, then text and non-textual content can be identified, but computational efficiency decreases
Solution Approach 1:
The patent merges traditional OCR text recognition and object recognition for non-textual elements into a single machine learning model. This unified approach shares computational resources and processing pathways between what were previously separate tasks, reducing overall computational resource usage while maintaining the ability to identify both text and non-textual document content.
Data Source
AI summary
Aspects and implementations provide for techniques of fast and efficient identification of objects of multiple types in electronic documents. The disclosed techniques include, for example, processing, using a machine learning model (MLM), an image of a document to generate a plurality of pixel-level maps (PLMs), characterizing associations of pixels of the image with various object types. The MLM includes a backbone neural network (NN) processing the image and generating a feature tensor for the image. The MLM further includes a plurality of classification NNs that process the feature tensor and generate PLMs. The techniques further include generating, using the PLMs, an object-level map identifying placement of one or more objects in the document. The classification NNs may be trained together (end-to-end) with the backbone NN.