Satellite Image AI Explainability With Local Attention Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing visual explanation techniques for satellite images using convolutional neural networks (CNNs) suffer from background bias due to pooling layers, which lose small object information and reduce recognition rates, necessitating improved methods to enhance explainability and accuracy.
Innovation Solution
An electronic device employs a two-stage AI model comprising convolution blocks with pooling layers followed by local attention blocks to amplify feature maps, emphasizing relevant regions, thereby improving object recognition and reducing background bias.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If pooling layers are used in CNN for computational efficiency, then processing speed is improved, but small object information is lost due to downsampling
Solution Approach 1:
The patent divides the feature extraction process into multiple parallel pathways: a main CNN pathway with pooling layers for efficient processing, and separate pathways that extract features at multiple intermediate stages before pooling. This segmentation allows the system to maintain computational efficiency while preserving fine-grained spatial information through the parallel non-pooled pathways.
Solution Approach 2:
The patent implements a nested structure where multiple feature maps at different hierarchical levels are integrated. The attention mechanism embeds fine-grained features from shallow layers within the coarser features from deep layers, creating a nested representation that preserves small object information while maintaining the computational efficiency of pooled features.
2Device complexity
If visual explanation is generated using single top-level convolution layer features, then computational complexity is reduced, but background bias occurs and recognition rate decreases
Solution Approach 1:
The patent applies local quality by generating attention maps at multiple hierarchical levels with different spatial resolutions. Shallow attention maps capture local fine-grained details, while deeper attention maps capture global contextual information. This multi-level approach ensures that visual explanations reflect both local object characteristics and global scene context, eliminating background bias without excessive computational complexity.
Solution Approach 2:
The patent adds a hierarchical dimension to visual explanation generation by incorporating features from multiple convolutional layers rather than relying solely on the top-level features. This dimensional expansion allows the system to capture both fine-grained and coarse-grained information, improving recognition accuracy while maintaining reasonable computational complexity through efficient feature fusion.
3Measurement precision
If attention maps from multiple layers are integrated, then object recognition accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent introduces an attention mechanism as an intermediary that efficiently integrates features from multiple layers. Instead of directly combining all features from every layer, the attention mechanism selectively weights and fuses features based on their relevance, reducing the computational burden of multi-level feature integration while maintaining high object recognition accuracy.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An electronic device includes a memory configured to store at least one instruction; and at least one processor configured to execute the at least one instruction to: input first data to a first artificial intelligence model including a plurality of convolution blocks sequentially connected with a pooling layer interposed therebetween to obtain a plurality of feature maps that are output by corresponding ones of the plurality of convolution blocks, input the first data and the plurality of feature maps to a second artificial intelligence model including a plurality of local attention blocks sequentially connected to obtain a plurality of attention maps that are output by corresponding ones of the plurality of local attention blocks, output an amplified feature map by amplifying a region corresponding to a last attention map among the plurality of attention maps in a last feature map among the plurality of feature maps, and input the amplified feature map to a classifier to output a classification result for the first data.