Data augmentation using image translation for underwater sonar image segmentation.

Area of Science:

• Robotics
• Computer Vision
• Marine Technology

Background:

• Object recognition is crucial for underwater unmanned vessels.
• Acquiring extensive experimental data for deep learning is challenging due to resource limitations.
• Current methods require significant effort in preparing real sonar image datasets.

Purpose of the Study:

• To generate synthetic sonar images for training deep learning models.
• To reduce the burden of manual data preparation for underwater object recognition.
• To evaluate the effectiveness of synthetic data in improving pixel segmentation performance.

Main Methods:

• Utilized the Pix2Pix image transformation model to synthesize sonar images.
• Applied generated synthetic data alongside real data to train a Fully Convolutional Network (FCN) model.
• Evaluated model performance using mean accuracy and mean Intersection over Union (IoU) metrics.

Main Results:

• Training with both synthetic and real data resulted in a mean accuracy of 0.81, a 6% increase compared to using only real data (0.7525).
• The mean IoU remained similar, showing 0.7225 with combined data versus 0.7275 with real data only.
• Pix2Pix successfully generated realistic sonar image-mask pairs, simplifying the training data preparation process.

Conclusions:

• Synthetic data generation using Pix2Pix is a viable approach to augment limited real-world underwater datasets.
• The proposed method effectively enhances the performance of deep learning models for underwater object recognition.
• This technique alleviates the significant effort required for manual annotation of sonar imagery.