USV Environment Perception With Brain-Like Memory for Sea Decisions

The patent introduces an image feature extractor as an intermediary component that processes visual environmental information through multiple layers (convolutional layers, pooling layers) to extract meaningful features. This intermediary processing stage transforms raw image data into structured feature representations that are more reliable for decision-making under complex sea conditions, bridging the gap between automated input and accurate environmental perception.

The patent replaces traditional mechanical/sensor-based environmental information acquisition with a vision-based computational system. Instead of relying on physical sensors that struggle in complex sea conditions, the system uses image processing and deep learning models (including BERT and large language models) to extract and interpret environmental information, substituting physical measurement with computational analysis.

The patent segments the decision-making system into distinct functional modules: image feature extractor, BERT model for text feature extraction, fully connected layers, and large language model. Each module handles a specific aspect of processing (visual features, textual features, integration, and decision generation), allowing the complex system to be managed through modular components that can be independently optimized and trained.

The patent employs a large language model that serves multiple functions: it processes textual environmental descriptions, integrates visual features from the image extractor, performs reasoning about sea conditions, and generates decision outputs. This multi-functional approach consolidates what would otherwise require separate specialized models, managing complexity while maintaining high decision accuracy through a single versatile model.