Image Labeling via Partial Feature Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image labeling methods are inefficient in generating accurate labels for images, particularly for painting images, as they rely solely on global features, leading to low accuracy in categorization and recognition due to the lack of detailed information.
Innovation Solution
A method that acquires both global and partial image features by using a neural network model with feature extracting networks and a target feature fusing network to generate multiple labels, including category, subject, and content labels, by fusing features of different dimensions and utilizing recurrent neural networks for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If only global features are used for image labeling, then the labeling process is simple and fast, but the accuracy of category, subject, and content label generation is low
Solution Approach 1:
The patent divides the image into multiple partial images at different resolutions (low-resolution partial images and high-resolution partial images) and extracts features separately from each segment. This segmentation allows the system to capture both global context and detailed local information, resolving the contradiction between labeling accuracy and system complexity by organizing the complex task into manageable segments.
Solution Approach 2:
The patent introduces a multi-dimensional feature extraction approach by extracting features at different resolution levels (low-resolution and high-resolution) and combining them through a fusion network. This dimensional transformation enables the system to leverage both coarse global features and fine detailed features, significantly improving labeling accuracy without overwhelming complexity.
2Measurement precision
If partial images are extracted and processed to improve labeling accuracy, then the recognition of detailed features is enhanced, but the processing time and computational resources increase
Solution Approach 1:
The patent extracts and processes only the necessary partial images at different resolutions rather than processing the entire image at all resolutions. By selectively extracting partial regions and using a fusion network to combine only the essential features, the system achieves improved recognition accuracy while reducing unnecessary computational time and resources.
Solution Approach 2:
The patent performs preliminary feature extraction on partial images at different resolutions before final label generation. The low-resolution partial images provide a quick initial feature set, and only these pre-extracted features are fed into the fusion network, reducing overall processing time while maintaining detailed feature recognition capability.
3Reliability
If a neural network model is trained with extensive painting images to improve labeling accuracy, then the model achieves high performance, but the training data requirement and training time increase significantly
Solution Approach 1:
The patent changes the approach from training with extensive painting images to training with non-painting images and then adapting the model. By pre-training on non-painting images and using a transfer learning approach with feature fusion, the model achieves high reliability for painting image labeling without requiring large quantities of painting-specific training data, significantly reducing training data requirements.
Solution Approach 2:
The patent introduces non-painting images as an intermediary training resource. Instead of directly training on scarce painting images, the system pre-trains on abundant non-painting images and then adapts to painting images through the feature fusion mechanism, achieving reliable performance with minimal painting image training data.
4Adaptability or versatility
If multiple features from different resolutions are fused to generate comprehensive labels, then the completeness of category, subject, and content labels is improved, but the system complexity and feature processing burden increase
Solution Approach 1:
The patent merges low-resolution partial image features and high-resolution partial image features through a fusion network to generate comprehensive labels. By combining features from different resolution levels in a structured manner, the system achieves complete labeling (category, subject, and content) while managing complexity through the organized integration of multiple feature sources.
Data Source
AI summary
Provided is a method for generating an image label, including: acquiring a partial image of a target image after acquiring the target image with a label to be generated; then, acquiring a plurality of features based on the target image and the partial image, wherein the plurality of features include a first feature of the target image and a second feature of the partial image; and finally, generating a first-type image label of the target image based on the first feature and the second feature.