Semantic Picture Coding With CNN Features and Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing picture coding technologies fail to detect the most meaningful parts of a picture and lack connection with high-level semantics, leading to poor recognition capabilities and inability to effectively identify plagiarism in scenarios like e-commerce.
Innovation Solution
A method involving training an initial feature extraction model with CNN to obtain a target feature extraction model, followed by training an initial clustering model with PQ, and performing picture coding using both models to achieve semantic-level feature extraction and dimension reduction, enhancing recognition capabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Stability of the object's composition
If existing picture coding technology captures picture features at pixel level, then geometric deformation resistance is improved, but recognition capability deteriorates
Solution Approach 1:
The patent transitions from pixel-level feature extraction to semantic-level feature extraction by introducing deep learning models (CNN, RNN, Transformer) that operate at multiple abstraction layers. This dimensional shift enables the system to capture both geometric properties and semantic meaning simultaneously, resolving the contradiction between geometric deformation resistance and recognition capability.
Solution Approach 2:
The patent combines multiple feature types (geometric features, semantic features, texture features) into a composite feature representation. By integrating features from different sources and levels through ensemble methods or multi-modal fusion, the system achieves both geometric stability and semantic recognition accuracy.
2Measurement precision
If existing picture coding technology uses pixel-level features, then edge information capture is improved, but connection with high-level semantics deteriorates
Solution Approach 1:
The patent segments the feature extraction process into multiple stages: low-level feature extraction (edges, textures) followed by high-level semantic feature extraction through deep neural networks. This segmentation allows preservation of precise edge information while progressively building up to semantic understanding, preventing information loss at any single stage.
Solution Approach 2:
The patent implements a hierarchical feature extraction architecture where low-level features (edges, contours) are nested within mid-level features (shapes, objects), which are in turn nested within high-level semantic features (scene understanding, concept recognition). This nested structure preserves fine-grained edge information while integrating it into broader semantic contexts.
Data Source
AI summary
Embodiments of the application disclose a method and device for picture coding, and a storage medium, the device for picture coding is configured to train an initial feature extraction model with first training data, to obtain a target feature extraction model; train an initial clustering model with second training data, to obtain a first clustering model; and perform picture coding on picture data according to the target feature extraction model and the first clustering model, to obtain coded data corresponding to the picture data.