Attribute-Guided Image Classification for Robust Domain Shift

The patent introduces attribute predictions as an intermediary between image features and class classification. The network first predicts attributes (e.g., 'has wings', 'is red') that describe properties of the image, then uses these attributes to inform the final class prediction. This intermediary layer provides explainability by showing which attributes contributed to the classification, while maintaining automation throughout the process.

The patent segments the classification task into two distinct parts: attribute prediction and class prediction. The network outputs both attribute predictions (providing explanation) and class predictions (providing classification). This segmentation allows the system to maintain relatively simple architecture while generating explanatory attributes that localize why certain predictions were made, addressing the explainability gap without excessive complexity.

The patent performs preliminary attribute prediction before final class classification. By first predicting attributes that capture essential properties of the image (such as color, shape, presence of specific features), the network prepares a more robust representation that can better handle domain shifts. This preliminary action of attribute extraction creates a bridge between different domains while maintaining classification speed through efficient network architecture.