Adaptive Feature Vector Sizing for Scalable Document Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing document classification methods using neural networks face challenges when the number of document types to be classified increases, requiring relearning of the entire model and often result in insufficient feature vector dimensions for accurate classification.
Innovation Solution
A method and device that adaptively change the dimension of feature vectors by predicting and modifying the feature extraction model to ensure sufficient dimensions for classification, using a dimension prediction model that can learn without labeled data and adjust the feature extraction model to expand dimensions as needed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a fully connected layer is coupled to a feature extraction model to classify N document types, then accurate classification of N document types can be achieved, but the entire neural network model needs to be relearned when document types increase
Solution Approach 1:
The system segments the classification task into two independent parts: a feature extraction model that generates fixed-dimensional feature vectors, and a clustering model that performs classification based on these vectors. This segmentation allows the feature extraction model to remain unchanged when document types increase, eliminating the need to relearn the entire network while maintaining classification accuracy through the adaptive clustering component.
Solution Approach 2:
The system introduces dynamic adaptability by using clustering models (such as k-mean clustering) that can dynamically adjust to varying numbers of document types without requiring retraining of the feature extraction model. The clustering algorithm adapts its parameters based on the actual number of document types present, enabling the system to handle dynamic changes in classification categories efficiently.
2Device complexity
If feature vectors of fixed dimension are used for document classification, then the feature extraction model structure remains simple, but feature vectors may fail to provide sufficient information when the number of document clusters increases
Solution Approach 1:
The system addresses information sufficiency by extending the feature vector dimensionality through fully connected layers that transform fixed-dimensional features into expanded feature representations. This dimensional transformation enables the feature vectors to capture sufficient information for clustering a large number of document types while maintaining a relatively simple base feature extraction model structure.
Solution Approach 2:
The system introduces fully connected layers as intermediary components between the feature extraction model and the clustering model. These intermediary layers serve to transform and enrich the feature representations, providing sufficient information for accurate classification of numerous document types without requiring complex modifications to the original feature extraction model.
3Measurement precision
If the dimension of feature vectors is increased to provide sufficient information for more document clusters, then classification accuracy can be maintained, but the complexity and computational cost of the feature extraction model increases
Solution Approach 1:
The system segments the complexity into two parts: a simple, fixed-structure feature extraction model that generates initial feature vectors, and a separate dimensionality expansion component (fully connected layers) that adapts the feature dimensions as needed. This segmentation allows the core feature extraction model to remain simple while still achieving high classification accuracy through adaptive dimension expansion in the subsequent processing stages.
Data Source
AI summary
The present disclosure relates to a document classification method and a document classification device. A document classification method performed by a processor inside a computing device according to an embodiment of the present disclosure may include: obtaining a predicted dimension which is a dimension of a feature vector necessary to classify a document image that has been input, through a learned dimension prediction model; generating a feature vector of the document image, through a feature extraction model based on the predicted dimension; and identifying a document type corresponding to the document image on the basis of the generated feature vector.