Adversarial Neural Network for Robust Target Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for target recognition using deep neural networks (DNNs) face challenges in extracting robust features due to biased features, which can lead to incorrect recognition results, especially when trained with data from a single domain, and are costly in terms of data collection and processing.
Innovation Solution
A learning apparatus comprising multiple neural networks, including a first neural network for feature extraction, a second neural network with a different structure for biased feature extraction, and a learning support neural network that trains to reduce biased features, allowing for adaptive extraction of robust features by adversarial learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If learning is performed using a wide variety and number of data sets from different domains, then robustness of target recognition is improved, but data collection cost and processing cost increase enormously
Solution Approach 1:
The patent extracts and removes biased features from the feature vector using the learning support neural network. By identifying and eliminating the biased feature component that causes domain-specific overfitting, the system achieves robust target recognition without requiring extensive multi-domain training data, thus reducing data collection and processing costs while maintaining reliability.
Solution Approach 2:
The learning support neural network acts as an intermediary component that bridges the main neural network and the biased feature removal process. It receives the feature vector, identifies biased features through its training objective, and outputs a debiased feature representation, enabling the system to achieve domain robustness without direct access to multiple domain data sets.
2Loss of information
If learning is performed by considering biased features in addition to features to be noticed, then the model captures more information, but correct recognition results cannot be output due to influence of biased features
Solution Approach 1:
The patent segments the feature extraction process into two distinct components: the main neural network that extracts comprehensive features including biased ones, and the learning support neural network that specifically identifies and removes biased features. This segmentation allows the system to first capture all relevant information and then precisely eliminate the harmful biased components, achieving both complete information capture and high recognition accuracy.
Solution Approach 2:
The patent converts the harmful biased features into a beneficial training signal. By using the biased features as the target for the learning support neural network to predict and remove, the system transforms the problematic biased information into a useful component for learning what to eliminate, thereby improving recognition accuracy while maintaining comprehensive feature capture.
3Ease of manufacture
If a specific model for extracting texture feature is specified based on assumption that biased feature is texture feature, then extraction of biased feature is facilitated, but the technique is limited to cases where texture feature is treated as biased feature
Solution Approach 1:
The patent creates a universal learning support neural network that can handle any type of biased feature, not just texture features. The network uses a generic training objective based on domain invariance that adapts to whatever biased features are present in the data, making the technique applicable to various domains and bias types without requiring domain-specific model design, thus achieving both ease of implementation and broad versatility.
Data Source
AI summary
A learning apparatus executes processing of: a first neural network that extracts a first feature of a target in image data; a second neural network that extracts a second feature of the target in the image data using a network structure different from the first neural network; and a learning support neural network that extracts a third feature from the first feature extracted by the first neural network. Here, the second feature and the third feature are biased features for the target. The learning apparatus trains the learning support neural network so that the second feature and the third feature come closer, and trains the first neural network so that the third feature appearing in the first feature extracted by the first neural network is reduced.