Machine Learning Model for Adaptive Image Region Selection

The patent segments the single machine learning model into multiple specialized models, each trained to evaluate images for specific usage purposes (e.g., landscape, portrait, fashion). Each model outputs an attention map focused on its specialized domain, and these attention maps are combined to produce the final cropping result. This segmentation allows the system to maintain low device complexity while achieving high adaptability across different image usage contexts.

The patent creates a universal machine learning framework that can be applied across multiple image usage purposes. By training specialized models within this universal framework, the system achieves multi-functionality where the same overall architecture and training methodology can be reused for different purposes (landscape, portrait, fashion, etc.). This universality reduces the learning data preparation workload by allowing the system to leverage the same training pipeline and architectural patterns across different specialized models, rather than creating entirely separate systems for each purpose.

The patent merges the attention maps produced by multiple specialized machine learning models into a single composite attention map. This merging process combines the strengths of each specialized model's attention focus, resulting in a more precise identification of the principal portion that accounts for the specific image usage purpose. The merging operation integrates the different attention perspectives (landscape, portrait, fashion) to produce a comprehensive and accurate cropping region, improving measurement precision while maintaining manageable device complexity through efficient combination algorithms.