Training Data Generation for Low-Frequency Class Recognition

The patent applies local quality by selecting specific classes for data mixing based on their recognition accuracy performance. Instead of uniformly mixing all classes, it identifies low-frequency classes (those with poor recognition accuracy) and selectively mixes their training data with other classes to improve their specific performance while leaving other classes unchanged.

The patent changes the parameter of class selection from random to based-on-recognition-accuracy. By calculating recognition accuracy for each class and selecting low-frequency classes accordingly, the data augmentation process is directed toward classes that need improvement, transforming the uniform random approach into a targeted selective approach.

The patent creates synthetic training data by copying and mixing existing image data from different classes. Instead of collecting new real-world data for low-frequency classes, it generates artificial training samples by combining and transforming existing data, thereby avoiding the high costs associated with data collection while still providing sufficient training material.

The patent uses data augmentation as an intermediary process to bridge the gap between limited available data and the need for more training samples. By introducing a data mixing mechanism that operates on existing data, it creates additional training samples without requiring direct data collection, thus reducing costs while improving model performance.