Mini-batch Learning Apparatus for Semantic Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing mini-batch learning methods for semantic segmentation often result in low rare class determination accuracy due to class bias in learning input and annotation images, leading to inadequate model performance.
Innovation Solution
A mini-batch learning apparatus that calculates area ratios of classes, specifies correction target classes based on these ratios, and adjusts loss values using a loss function to balance the representation of rare and non-rare classes, thereby enhancing the model's determination accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If mini-batch learning is performed using conventional methods without class balance consideration, then learning speed and processing efficiency are maintained, but rare class determination accuracy deteriorates due to class bias in the data
Solution Approach 1:
The patent applies preliminary action by calculating area ratios and identifying correction target classes before the actual mini-batch learning process. The loss value correction is prepared in advance based on class distribution analysis, ensuring rare classes are properly weighted from the start of learning without adding complexity during the learning execution phase.
Solution Approach 2:
The patent changes the loss value parameter dynamically based on class distribution. By adjusting loss values for correction target classes (typically rare classes with low area ratios) relative to other classes, the learning process adapts to class imbalance without requiring complex architectural modifications to the model or data sampling strategies.
2Measurement precision
If loss values are corrected for all classes uniformly, then class balance is improved, but learning efficiency and convergence speed deteriorate due to unnecessary processing overhead
Solution Approach 1:
The patent applies local quality by selectively correcting loss values only for specific correction target classes (typically rare classes) rather than uniformly adjusting all classes. The area ratio calculation identifies which classes need correction, and only those classes have their loss values adjusted, preserving learning efficiency while improving accuracy for underrepresented classes.
Solution Approach 2:
The patent uses partial action by applying loss value correction only to the extent necessary - specifically to correction target classes identified by low area ratios. This avoids excessive correction of already well-represented classes, maintaining learning efficiency while sufficiently addressing the class imbalance problem.
3Measurement precision
If area ratio calculation and loss value correction are implemented, then rare class representation is improved, but computational overhead and processing time increase
Solution Approach 1:
The area ratio calculation and correction target class identification are performed as preliminary steps before mini-batch learning begins. This upfront preparation ensures that the actual learning process uses pre-computed correction factors, avoiding repeated calculations during training and minimizing the time penalty to just the initial setup phase.
Solution Approach 2:
The patent replaces complex mechanical data sampling or data augmentation mechanisms with a simpler computational approach - adjusting loss values through mathematical operations. This substitution maintains accuracy improvement while avoiding the computational overhead of generating additional data samples or complex resampling procedures during training.
Data Source
AI summary
In a case where the operation program is started, a CPU of the mini-batch learning apparatus functions as a calculation unit, a specifying unit, and an evaluation unit. The calculation unit calculates an area ratio of each of a plurality of classes in mini-batch data. The specifying unit specifies, as a correction target class, a rare class of which the area ratio is lower than a setting value. The evaluation unit evaluates the class determination accuracy of the machine learning model by using a loss function. As correction processing, the evaluation unit sets a weight for a loss value of the rare class to be larger than a weight for a loss value of a class other than the rare class.