Sonar Object Detection via Pseudo-Optical Translation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Interpreting sonar imagery requires trained personnel and distinguishing objects of interest from background in echograms is challenging due to the dissimilarity between sonar and electro-optical images, and the limited availability of annotated echogram data for training models.
Innovation Solution
A deep neural network with feature extraction layers trained using non-sonar image data and classification layers trained using sonar image data is employed to detect and classify objects in echograms, enabling automated detection and classification without human intervention, and associating geolocation information with detected objects using bathymetry data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If echogram images are used for underwater object detection, then detection capability in underwater environments is achieved, but interpretation difficulty increases due to dissimilarity between sonar and electro-optical images
Solution Approach 1:
The patent introduces an intermediary translation layer that converts sonar echogram images into pseudo-electro-optical images that resemble conventional optical images. This intermediary representation allows standard electro-optical image processing algorithms to be applied, thereby reducing interpretation difficulty while maintaining underwater detection capability. The translation process creates a bridge between the dissimilar sonar and electro-optical image domains.
2Measurement precision
If trained personnel are used to interpret sonar imagery, then detection accuracy improves, but operational complexity and time consumption increase
Solution Approach 1:
The system employs automated machine learning algorithms that perform object detection and classification without requiring human operators to manually interpret sonar images. The trained neural networks automatically analyze echograms, identify objects of interest, and classify them, thereby maintaining high detection accuracy while eliminating the need for specialized human expertise and reducing operational complexity.
3Measurement precision
If annotated echogram data is used for training models, then detection accuracy improves, but data availability decreases due to limited annotated datasets
Solution Approach 1:
The patent applies transfer learning by pre-training feature extraction models on large datasets of electro-optical images before fine-tuning them on smaller sonar echogram datasets. This preliminary training on abundant optical images allows the model to learn general image features that can be transferred to the sonar domain, thereby achieving good detection accuracy even with limited annotated echogram data available for training.
Data Source
AI summary
Surveillance systems and methods taught herein provide automated detection and classification of objects of interest in a submerged or underwater environment such as a body of water. The sonar systems and methods taught herein can detect and classify a variety of objects in echograms without feedback or instructions from a human operator. The sonar systems and methods taught herein include a data model that is partially trained using non-echogram image data and can associate geolocation information with detected objects of interest.