Attention-Based Deep Learning Networks Could Improve Sonar Systems

Researchers in China and the United States recently explored how an attention-based deep neural network (ABNN) could help improve sonar systems. 

The research was published in the Journal of the Acoustical Society of America by the Acoustical Society of America through AIP Publishing. 

Qunyan Ren is co-author of the research. 

“We found the ABNN was highly accurate in target recognition, exceeding a conventional deep neural network, particularly when using limited single-target data to detect multiple targets,” Ren said.

DNNs and ABNNs

Deep learning, which is a machine-learning method that uses artificial neural networks that work to recognize patterns, relies on layers of artificial neurons (nodes) that learn a distinct set of features based on the information present in the previous layer. 

Attention-based deep neural networks use an attention module to mimic certain elements in the cognitive process in humans. These elements specifically help us focus on the most important parts of language, an image, or some other pattern while tuning out the rest.

ABBNs achieve this by adding more weight to certain nodes, which enhances specific pattern elements in the machine-learning process.

Incorporating ABNN Into Sonar

By incorporating an ABNN system into sonar equipment for targeted ship detection, the team of researchers was able to test two ships in a 135-square-mile, shallow area in the South China Sea. The results were compared with a regular deep neural network (DNN), and other equipment like radar helped determine over 17 interfering vessels in the area that was tested. 

The researchers found that the ABNN increases its predictions as it moves toward the features that closely correlate with the training goals. As the network continually cycles through the training dataset, detection becomes more pronounced. This accentuates the weighted nodes and disregards irrelevant data.

The ABNN accuracy of detecting ships A and B separately was slightly higher than the DNN, with the former achieving 98% and the latter 97.4%. When it came to the ABNN accuracy of detecting both ships in the same vicinity, it was even higher at 74%, compared to the DNN’s 58.4%.

A traditional ABNN model is usually trained with multiship data if it is being used for multiple-target identification. However, this process can quickly become costly and complex. Because of this, the researchers trained the ABNN model to detect each target separately. As the output layer of the network is extended, the individual-target datasets merge.

“The need to detect multiple ships at one time is a common scenario, and our model significantly exceeds DNN in detecting two ships in the same vicinity,” Ren said. “Moreover, our ABNN focused on the inherent features of the two ships simultaneously.”