Estimation Model Training With Area And Region Relationships
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition technologies face challenges in training estimation models for accurately segmenting multiple regions in images, particularly in medical imaging, due to the time and skill required for adding area information as correct answer masks, making it difficult to prepare sufficient training data.
Innovation Solution
An information processing apparatus and method that utilizes both first training data with area information and second training data with relationship information to train an estimation model, using evaluation values to reduce loss and improve segmentation accuracy, even with limited training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If area information (correct answer mask) is added to training images to train the estimation model, then segmentation accuracy is improved, but the time and skill required for data preparation increases significantly
Solution Approach 1:
The patent introduces relationship information as an intermediary element that bridges the gap between limited annotated data and accurate segmentation. Relationship information (inclusion relationships between regions) serves as a mediator that provides additional structural constraints without requiring extensive pixel-level annotations, thus reducing preparation time while maintaining segmentation accuracy
Solution Approach 2:
The patent segments the training data into two distinct components: area information (traditional pixel-level masks) and relationship information (inclusion relationships between regions). This segmentation allows the system to leverage both types of data independently, reducing the burden of complete pixel-level annotation while preserving segmentation accuracy through the complementary relationship constraints
2Reliability
If area information is added to all training images, then the estimation model can be trained effectively, but the complexity of data preparation and processing increases
Solution Approach 1:
The patent divides training data into two separate categories (area information and relationship information) that can be processed and stored independently. This segmentation reduces the complexity of data preparation by allowing different processing pipelines for each type, while both contribute to reliable model training through their complementary nature
Solution Approach 2:
Relationship information serves multiple functions simultaneously: it provides structural constraints for segmentation, reduces the need for extensive pixel-level annotations, and offers an alternative supervision signal for regions where area information is unavailable or ambiguous, thereby improving training effectiveness without proportionally increasing complexity
3Measurement precision
If relationship information is incorporated as a restriction condition, then segmentation accuracy with limited data is improved, but the computational complexity of training increases
Solution Approach 1:
The patent transforms the training objective by incorporating relationship information as additional constraint parameters in the loss function. By changing the parameter space to include both area information and relationship information, the model achieves better segmentation accuracy with limited data while the computational overhead is managed through efficient gradient computation for the additional constraints
Data Source
AI summary
A processor acquires a plurality of first training data in which area information indicating an area in which each of a plurality of regions is present is added to a first training image which is at least a part of a plurality of training images each including the plurality of regions, and a plurality of second training data in which relationship information indicating a relationship between the plurality of regions is added to a second training image which is at least a part of the plurality of training images. The processor calculates, for each first training image, a first evaluation value for training an estimation model such that the plurality of regions specified by using the estimation model match the area information. The processor derives, for each second training image, estimation information in which the relationship indicated by the relationship information is estimated by using the estimation model to calculate a second evaluation value indicating a degree of deviation between the estimation information and the relationship information. The processor trains the estimation model such that a loss including, as elements, the first evaluation value and the second evaluation value is reduced.