CNN Feature Representation via Agreement Rate Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current CNN-based image identification technologies face challenges in efficiently representing features captured by specific layers, making it difficult to analyze the factors leading to identification results, particularly when errors occur, and require extensive human effort for visualization.
Innovation Solution
A feature representation device and method that calculates agreement rates for feature maps across predefined concepts, clusters feature maps, integrates them using agreement rates, and generates visualization images with threshold values to efficiently represent features captured by CNNs, allowing for the analysis of identification results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If CNN-based image identification technology is used, then identification accuracy is improved, but it becomes difficult to analyze factors leading to identification results
Solution Approach 1:
The patent segments the CNN feature extraction process by isolating feature maps from specific layers and filters, allowing individual feature analysis. This segmentation enables the system to examine intermediate representations without disrupting the overall identification accuracy, thus resolving the contradiction between maintaining high accuracy and enabling feature analysis.
Solution Approach 2:
The patent introduces an intermediary visualization system that bridges the gap between CNN internal features and human interpretable representations. By creating visual mappings of feature maps to original image regions, this intermediary layer preserves identification accuracy while making features analyzable, thus resolving the information loss problem.
2Loss of information
If feature maps from all filters are visualized, then comprehensive feature analysis is achieved, but extensive human effort and time are required
Solution Approach 1:
The patent extracts and prioritizes only the most relevant filters and feature maps based on their contribution to identification results. By selecting top-contributing features rather than visualizing all filters, the system maintains comprehensive feature coverage while dramatically reducing visualization time and human effort required.
Solution Approach 2:
The patent applies partial action by visualizing a representative subset of feature maps rather than all possible filters. This selective approach provides sufficient feature coverage for analysis while avoiding the excessive time investment required for complete visualization of all CNN filters.
3Reliability
If detailed feature analysis is enabled, then error cause analysis is improved, but device complexity increases
Solution Approach 1:
The patent performs preliminary organization and classification of feature maps before detailed analysis is needed. By pre-processing and structuring feature representations in an analyzable format, the system enables detailed error analysis when needed without adding complexity to the core CNN identification process.
Data Source
AI summary
The present invention relates to representing image features used by a convolutional neural network (CNN) to identify concepts in an input image. The CNN includes a plurality of filters in each of a plurality of layers. The method generates the CNN based on a set of images for training with predetermined concepts in regions of the set of images. For a select layer of the CNN, the method generates integrated maps, Each integrated map is based on a set of feature maps in a cluster and relevance between the set of feature maps for the select layer and a region representing one of the features in the image data. The method provides a pair of a feature representation visualization image of a feature in the select layer and a concept information associated with the integration map.