Sketch Classification Training Using Real-Image Feature Transfer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing classification methods for sketch recognition, such as using support vector machines or classification networks, face issues of overfitting or underfitting due to scarce training samples, affecting the accuracy of identifying hand sketches.
Innovation Solution
A classification training method that utilizes both sketch and real graph classification models, incorporating feature extraction and classification modules, adjusts fixed parameters using loss functions to enhance accuracy by leveraging information from both models during training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional classification methods are used with limited training samples, then the training process is simpler, but overfitting or underfitting occurs reducing classification accuracy
Solution Approach 1:
The patent combines sketch classification model and real graph classification model into a unified training framework. The loss function integrates both classification losses, allowing the models to learn complementary features. This merging enables the system to leverage diverse data sources (sketches and real graphs) to improve generalization capability while maintaining classification accuracy, directly resolving the overfitting/underfitting problem caused by limited training samples.
Solution Approach 2:
The patent introduces a feature analysis model as an intermediary component that analyzes features extracted by both the sketch feature extraction module and real graph feature extraction module. This intermediary processes and integrates features from different sources before classification, enabling better feature representation and improving model reliability without requiring extensive training data.
2Measurement precision
If only sketch classification model is trained, then the training process is faster, but the classification accuracy is insufficient
Solution Approach 1:
The patent performs preliminary feature extraction using dedicated feature extraction modules for both sketches and real graphs before the classification stage. The feature analysis model prepares and integrates these features in advance, creating a robust feature representation that improves subsequent classification accuracy. This preliminary processing enables more efficient training by reducing the computational burden during the classification optimization phase.
Solution Approach 2:
The patent segments the classification system into distinct functional modules: sketch feature extraction module, real graph feature extraction module, feature analysis model, sketch classification model, and real graph classification model. Each module has a specific function and can be trained independently or jointly, allowing flexible training strategies that balance accuracy and efficiency requirements.
3Measurement precision
If dual classification models are trained with integrated loss function, then classification accuracy improves, but the device complexity increases
Solution Approach 1:
The patent designs the feature analysis model to serve multiple functions: it analyzes features from both sketch and real graph extraction modules, integrates diverse feature types, and provides unified input to both classification models. This multi-functional component reduces the need for separate processing pipelines, managing system complexity while maintaining high classification accuracy through integrated learning.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A classification training method is provided, including: selecting a training set, determining a category of a sketch in the training set according to a sketch classification model, to obtain a first category processing result, and analyzing, according to a second feature analysis model, a feature of a sketch extracted by a first feature extracting model, to obtain an analysis result of a second sketch; then obtaining a function value of a first loss function according to the first category processing result and the analysis result of the second sketch; and finally adjusting a first fixed parameter value of the sketch classification model according to the function value of the first loss function.