Image Recognition Dataset Construction via Semantic Label Supplementation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The high cost and inefficiency of manual annotation processes hinder the establishment of large-scale multi-label image datasets, leading to information loss and confusion in deep learning models trained on single-label datasets.
Innovation Solution
A method and apparatus for establishing an image set using a network device that converts label content into word identifiers, constructs hierarchical semantic structures, and performs label supplementation based on semantic relationships and cooccurrence probabilities to create a target multi-label image set for training image recognition models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual annotation is used to create multi-label image datasets, then label accuracy can be maintained, but cost and time consumption increase significantly
Solution Approach 1:
The system enables automatic self-annotation by using a semantic network to automatically generate labels for images. The method retrieves semantic information from the semantic network based on image features and automatically creates multi-label annotations without requiring manual annotation for each image, thus achieving self-service annotation that reduces time consumption while maintaining accuracy through semantic relationship reasoning
Solution Approach 2:
The semantic network serves as an intermediary between image data and label data. Instead of directly annotating images manually, the system uses the semantic network as a mediator to retrieve relevant semantic information and automatically generate labels, thereby reducing the direct human effort required while maintaining label quality through semantic relationship inference
2Ease of manufacture
If single-label datasets are used for training, then data collection is simpler, but information loss occurs leading to model performance degradation
Solution Approach 1:
The method segments the label generation process into multiple stages: first retrieving semantic information from the semantic network, then generating candidate labels, and finally selecting appropriate labels based on image characteristics. This segmentation allows the system to systematically explore multiple semantic relationships and generate comprehensive multi-label annotations that capture richer information than single-label approaches
Solution Approach 2:
The system transitions from single-label to multi-label annotation by adding the dimension of semantic relationship reasoning. Instead of assigning only one label per image, the method leverages semantic networks to identify multiple related concepts and attributes, thereby enriching the label space and preventing information loss while maintaining relatively simple data collection processes
3Productivity
If automated label generation is implemented, then annotation efficiency improves, but label quality may deteriorate
Solution Approach 1:
The system incorporates feedback mechanisms where the generated labels are evaluated and refined based on semantic relationships and image characteristics. The method uses the semantic network to provide feedback on label relevance and adjusts the labeling process accordingly, ensuring that automated generation maintains high label quality through continuous semantic validation and refinement
Data Source
AI summary
A method of establishing an image set for image recognition includes obtaining a single-label image set comprising an image annotated with a single label, and a multi-label image set comprising an image annotated with a plurality of labels; converting content of each label into a corresponding word identifier according to a semantic network, to obtain a word identifier set, a converted single-label image set, and a converted multi-label image set; and constructing a hierarchical semantic structure according to the word identifier set and the semantic network. The method also includes performing label supplementation on the image in the converted single-label image set to obtain a supplemented single-label image set; performing label supplementation on the supplemented single-label image set to obtain a final supplemented image set; and establishing a target multi-label image set to train an image recognition model by using the target multi-label image set.