CNN Feature Map Clustering for Image Identification Explanation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image identification technologies, particularly those based on Convolutional Neural Networks (CNNs), struggle to provide detailed explanations of the features leading to identification results, making it difficult to analyze errors and interpret subtle differences in objects, such as material or shape variations.
Innovation Solution
A device and method that calculates feature maps from CNNs, clusters them, and assigns weights to explain the identification results by retrieving and mapping feature associations, allowing for detailed feature explanations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If CNN-based image identification is used to achieve high identification performance, then identification accuracy is improved, but the ability to analyze factors leading to identification results deteriorates
Solution Approach 1:
The patent segments the CNN's internal feature processing into individually accessible feature maps from different layers and filters. Each feature map is extracted and made available for separate analysis, allowing users to examine specific features detected by the CNN without losing the overall identification accuracy. This segmentation enables detailed examination of what features led to the identification result.
Solution Approach 2:
The patent introduces an intermediary system that bridges the CNN's internal feature processing and external interpretation. The system includes a feature map extraction unit that captures intermediate feature representations, a clustering unit that organizes these features, and a weight calculation unit that determines their contribution to the final result. This intermediary layer preserves both the CNN's high accuracy and provides interpretable feature information.
2Loss of information
If feature maps from CNN are extracted for analysis, then feature interpretability is improved, but device complexity increases
Solution Approach 1:
The patent designs a multi-functional system where the same CNN architecture serves both high-accuracy identification and feature interpretation purposes. The feature map extraction unit can operate at different CNN layers to provide features at multiple levels of abstraction, while the clustering and weight calculation units work together to organize and interpret these features. This universal approach avoids the need for separate interpretation models, reducing overall system complexity.
Solution Approach 2:
The patent performs preliminary organization of feature maps through clustering before detailed analysis. The clustering unit pre-groups feature maps based on their characteristics, creating a structured foundation for subsequent weight calculation and interpretation. This preliminary action simplifies the complexity of analyzing raw feature maps by organizing them into meaningful clusters first.
3Loss of information
If multiple feature maps from different CNN layers are analyzed, then detailed feature explanation is improved, but processing time increases
Solution Approach 1:
The patent applies local quality by allowing selective extraction and analysis of feature maps from specific CNN layers based on the analysis needs. Different layers provide different levels of feature detail - early layers capture low-level features while deeper layers capture high-level semantic features. The system can focus analysis on specific layers or clusters of layers, providing detailed feature explanations only where necessary, thus reducing overall processing time.
Solution Approach 2:
The patent implements partial action by enabling selective extraction of feature maps from only those CNN layers and filters that are most relevant to the specific identification task or error analysis need. Rather than processing all feature maps from all layers, the system can focus on a subset, reducing processing time while still providing sufficient detailed feature explanations for the given context.
Data Source
AI summary
An identification result explanation device calculates, for each of feature maps output individually from a plurality of filters used in specified layers of a CNN, a weight representing a degree of association with a result of identification for an input image by the CNN. The identification result explanation device outputs transposed feature maps obtained by transposing, for each of clusters, the feature maps in the cluster based on a result of classification of each of the feature maps for the input image classified into any of the clusters and on the weight calculated for each of the feature maps. The identification result explanation device uses the transposed feature maps in each of the clusters to retrieve, in a storage unit, each of the clusters including the feature maps linked to the same filters as those linked to the transposed feature maps for the input image and selects.