Neural Network Architecture for Incomplete Ground Truth Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing methods based on supervised learning, such as Fully Convolutional Networks, are unable to utilize training data when it includes missing ground truth data corresponding to at least one class of target objects, limiting their effectiveness in classification tasks like segmentation.
Innovation Solution
A neural network architecture comprising a common network shared across classes and single-class networks, where the common network extracts features and the single-class networks perform individual classifications, allowing training to proceed even with incomplete ground truth data by updating parameters only for classes with available data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a fully convolutional network is used for semantic segmentation, then segmentation capability is achieved, but training data with missing ground truth cannot be utilized
Solution Approach 1:
The network is segmented into a common network shared by all classes and multiple single-class networks, each responsible for one class. This segmentation allows the common network to learn from all available training data while single-class networks specialize in their respective classes, enabling utilization of training data with missing ground truth for the common network.
Solution Approach 2:
The common network serves multiple functions by being shared across all classes, extracting features that are useful for all target objects. This universal feature extraction capability allows the system to utilize training data even when ground truth is missing for some classes, as the common network can still learn from the available data.
2Measurement precision
If separate networks are used for each class, then classification accuracy is improved, but computational cost increases
Solution Approach 1:
The common network merges the feature extraction functionality that would otherwise be duplicated in separate networks for each class. By sharing this computational burden, the system reduces overall computational cost while maintaining the benefits of specialized single-class networks for accurate classification.
Solution Approach 2:
The common network performs universal feature extraction for all classes simultaneously, eliminating redundant computations that would occur if separate networks processed the same input image multiple times. This multi-functional approach reduces computational cost while preserving classification accuracy through specialized single-class networks.
3Measurement precision
If training data with missing ground truth is discarded, then training quality is maintained, but data utilization decreases
Solution Approach 1:
The training process is segmented into common network training using all available data and single-class network training using only complete data. This segmentation allows the system to utilize training data with missing ground truth for the common network while maintaining training quality for single-class networks through selective training.
Solution Approach 2:
The common network is trained with partial ground truth data (excluding classes with missing information), which is sufficient for learning general features. This partial action approach allows utilization of otherwise discarded data while maintaining training quality through subsequent specialized training of single-class networks.
Data Source
AI summary
A neural network training method for training a neural network that includes a common network that is common to a plurality of classes and a plurality of single-class networks that are connected to the common network and each having a single-class output layer for individually classifying each of the plurality of classes. An acquisition step acquires training data that includes ground truth data corresponding to at least one of the plurality of classes, and a training step that trains the neural network by using the training data. When training is performed by using training data that includes missing ground truth data corresponding to at least one of the plurality of classes, the training step trains a single-class network corresponding to a class of which ground truth data is not missing, and does not train the single-class network corresponding to the class of which ground truth data is missing.