Study on the Classification Performance of Underwater Sonar Image Classification Based on Convolutional Neural...

Area of Science:

• Robotics and Autonomous Systems
• Marine Technology
• Computer Vision

Background:

• Underwater human body detection is crucial but challenging due to poor visibility and sensor limitations.
• Sonar sensors are preferred for underwater imaging, but image quality is affected by background noise and incidence angles.
• Classifying sonar images requires addressing scattering and polarization noise for accurate detection.

Purpose of the Study:

• To develop an automated system for detecting submerged human bodies using sonar imagery.
• To improve the accuracy of sonar image classification in challenging underwater environments.
• To evaluate the effectiveness of Convolutional Neural Networks (CNNs) for this task.

Main Methods:

• Utilized sonar sensors for underwater image acquisition, overcoming limitations of vision sensors.
• Employed Convolutional Neural Networks (CNNs), specifically AlexNet and GoogleNet, for sonar image classification.
• Implemented data augmentation techniques focusing on scattering and polarization to enhance training data and model robustness.

Main Results:

• Achieved a high average classification accuracy of 91.6% using the GoogleNet model.
• Demonstrated the practicality of classifying sonar images even with data from simple testbed experiments.
• Successfully addressed noise issues in sonar images through CNNs and data augmentation.

Conclusions:

• Convolutional Neural Networks, particularly GoogleNet, are effective for automated underwater human body detection using sonar imagery.
• Data augmentation techniques are vital for improving CNN performance on noisy sonar data.
• The developed method shows promise for practical applications in underwater search and rescue.