Image Classifier Training With Domain-Invariant Feature Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The scarcity and high cost of manually labeled training images for image classifiers in automated driving systems, which limits their ability to generalize across different domains such as seasons, weather conditions, and sensor configurations, necessitating a method to effectively utilize training images from diverse domains without extensive labeling efforts.
Innovation Solution
A method that trains image classifiers to focus on domain-independent features by using a combination of image classifiers and domain classifiers, where training images from different domains are processed through convolution layers and domain classifiers to optimize classification scores and domain assignments, allowing the classifier to generalize better across domains without requiring extensive labeling, especially by utilizing radar and camera images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If training images from multiple domains are used to improve generalization, then the classifier's adaptability improves, but the complexity of training data collection and labeling increases
Solution Approach 1:
The patent uses domain adaptation techniques to create virtual copies of training data from one domain and transforms them to match target domains. By copying and adapting features from source domain training images, the system generates synthetic training examples for target domains without requiring actual physical data collection in those domains, thus improving generalization while avoiding the complexity of collecting and labeling data across all possible domains.
Solution Approach 2:
The patent employs domain adaptation by changing the parameters and features of training data representations. It learns domain-invariant features and transforms domain-specific features into a common representation space, effectively changing the parameterization of training data to make it applicable across multiple domains without requiring extensive domain-specific labeling efforts.
2Loss of time
If manual labeling of training images is reduced, then the cost and time of data preparation decreases, but the quality and accuracy of training may deteriorate
Solution Approach 1:
The patent implements self-supervised learning mechanisms where the system automatically generates labels and training signals from the data itself without requiring manual annotation. By using domain adaptation and feature alignment, the system creates its own training objectives and labels through unsupervised domain invariant feature learning, eliminating the need for extensive manual labeling while maintaining training quality through automated feature extraction and alignment procedures.
3Measurement precision
If domain-specific features are used for classification, then the accuracy on specific domains improves, but the ability to generalize to unseen domains deteriorates
Solution Approach 1:
The patent extracts and separates domain-invariant features from domain-specific features through domain adaptation techniques. By taking out the domain-specific components and focusing on learning domain-invariant representations, the system maintains classification accuracy on specific domains while simultaneously improving generalization to unseen domains. The domain adaptation module explicitly removes domain-specific biases from the feature representations.
Solution Approach 2:
The patent creates a universal feature representation that serves multiple domains simultaneously. Through domain adaptation and feature alignment, it learns a multi-functional feature extractor that can accurately classify objects across different domains using the same learned features. This universal representation achieves both domain-specific accuracy and cross-domain generalization by being trained to perform classification tasks across multiple domains with a single set of features.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Method (100) for training an image classifier (1) comprising the following steps: • training images (2a) are provided (110), wherein at least a first subset of the training images (2a) belongs to a first distribution (5a), at least a second subset of the training images belongs to a second distribution (5b), and these two distributions (5a, 5b) correspond to different domains of training images (2a); • target classification scores (4a) are provided (120) to which the image classifier (1) is to nominally map training images (2a) in the fully trained state; • the training images (2a) are supplied to the image classifier (1) as input images (2) and are mapped by the image classifier (1) to one or more classification scores (4) (130);• at least one output (11a-13a) of a convolution layer (11-13) resulting from this process is fed to a domain classifier (6) and mapped by this domain classifier (6) to domain assignments (7) of the training image (2a) to different domains (140), wherein this domain classifier can have a fully meshed layer (141) or generate a segmentation of the output (11a-13a) according to domains (143); • Parameters (1a) that characterize the behavior of the image classifier (1) are optimized (150) with the aim of minimizing the deviation of the classification scores (4) supplied by the image classifier (1) from the target classification scores (4a) during further processing of training images (2a), while maximizing the deviation of the domain assignments (7) from the actual domains of the respective training images (2a).