Spatially Invariant CNN Kernel Transfer Across Image Domains
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of transferring convolutional neural network (CNN) kernels between tasks with differing image qualities or color palettes is complicated by the need for manual selection and labeling of relevant weights, which is laborious and lacks user input in automatic methods, especially in safety-critical applications.
Innovation Solution
A method that clusters CNN kernels based on spatially invariant regions across datasets, allowing interactive selection and transfer of kernel weights through aggregation, resizing, and clustering, preserving interpretability and enabling user input.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual selection and labeling of kernel weights is performed, then transfer accuracy and interpretability are improved, but labor time and operational complexity increase
Solution Approach 1:
The system performs self-labeling of kernels by automatically generating labels based on clustering results and spatially invariant region analysis, eliminating the need for manual labeling while maintaining interpretability and transfer accuracy
Solution Approach 2:
Spatially invariant regions serve as an intermediary representation that bridges the source and target domains, enabling automatic kernel selection and transfer without manual intervention while preserving domain-specific important features
2Productivity
If automatic kernel transfer is performed, then productivity is improved, but user input and control are lost
Solution Approach 1:
The system provides feedback to users through visualizations of spatially invariant regions and kernel importance metrics, allowing users to review and adjust kernel selection while maintaining high automation efficiency
Solution Approach 2:
The kernel transfer process is made dynamic and adjustable, allowing users to interactively modify selection criteria and parameters based on feedback, balancing automation with user control
3Quantity of substance
If all kernels are transferred from source to target network, then completeness is improved, but computational complexity and processing time increase
Solution Approach 1:
The system extracts and transfers only the most relevant kernels identified through clustering and spatially invariant region analysis, rather than transferring all kernels, reducing processing time while maintaining transfer completeness for important features
Solution Approach 2:
Kernels are segmented into clusters based on their functional characteristics and spatial responses, allowing selective transfer of specific kernel groups rather than processing all kernels uniformly
4Measurement precision
If kernels are selected based on detailed feature analysis, then selection precision is improved, but device complexity and computational resources increase
Solution Approach 1:
The system performs detailed feature analysis only for a representative subset of kernels from each cluster rather than all kernels, achieving sufficient selection precision while reducing computational complexity through partial analysis
Data Source
AI summary
A computer-implemented method comprising: obtaining an output from each of a plurality of kernels in an extraction layer of a first trained convolutional neural network, wherein the first convolutional neural network is configured to identify one or more features in an image; aggregating the outputs corresponding to at least some input samples of a first domain to generate an aggregate map corresponding to that kernel; resizing the aggregate maps to a lower resolution to generate a plurality of region maps corresponding to the aggregate maps, respectively; clustering the region maps to generate clusters of region maps, each cluster comprising region maps having similar regions; and training, using input samples of a second domain, a second convolutional neural network with a kernel weight of at least one of the kernels which corresponds to at least one of the image regions of at least one of the clusters.