Classifier Algorithm Training via Image Grouping for Tagging

The patent segments the large set of training images into multiple subsets, each subset containing images with similar tags or semantic content. The classifier algorithm is trained on these smaller, manageable subsets separately, and then the results are aggregated. This segmentation approach maintains classification accuracy by ensuring comprehensive coverage of all tag categories while reducing the computational burden on any single training iteration, making the training process feasible within available processing power and storage constraints.

The patent performs preliminary organization of training images into subsets based on tag similarity or semantic content before training the classifier algorithm. This preliminary action ensures that even with limited computational resources, the training process systematically covers diverse semantic categories. By pre-organizing the data, the system maximizes the effectiveness of limited training resources while maintaining the ability to accurately classify various types of semantic content across multiple training passes.

The patent introduces a temporal dimension to the training process by iteratively training the classifier algorithm on multiple subsets of training images with different tag configurations. While each individual training pass may use a manageable number of tags, the aggregation of results across multiple passes enables the system to learn and maintain fine-grained classification capabilities for diverse categories. This multi-pass approach effectively adds a time dimension to the training process, allowing the system to handle complex classification tasks without requiring all tags to be processed simultaneously.