Method for Augmenting Side-Scan Sonar Seafloor Sediment Image Dataset Based on BCEL1-CBAM-INGAN
Haixing Xia1, Yang Cui1, Shaohua Jin1
1Department of Oceanography and Hydrography, Dalian Naval Academy, Dalian 116018, China.
Journal of Imaging
|September 27, 2024
Summary
This study introduces a new method using CBAM-BCEL1-INGAN to generate diverse seafloor sediment images, improving classification accuracy and addressing data scarcity for side-scan sonar analysis.
Area of Science:
- Marine geology
- Artificial intelligence
- Image processing
Background:
- Acquiring and labeling seafloor sediment image datasets is challenging.
- Existing datasets often lack sufficient diversity and quantity.
- Side-scan sonar technology requires high-quality, diverse data for accurate analysis.
Purpose of the Study:
- To propose a novel method for augmenting side-scan sonar seafloor sediment images.
- To address data scarcity and improve the diversity of training datasets.
- To enhance the performance of seafloor sediment classification models.
Main Methods:
- Integrated Convolutional Block Attention Module (CBAM) into INGAN generator for attribute learning.
- Introduced BCEL1 loss function (binary cross-entropy and L1 loss) in the discriminator for improved generation.
- Utilized AlexNet classifier to validate the authenticity and performance of augmented images.
Main Results:
- Demonstrated excellent performance in generating images of coarse sand, gravel, and bedrock.
- Achieved significant improvements in Frechet Inception Distance (FID) and Inception Score (IS).
- Increased bedrock recognition rate from 90.5% to 97.3% (6.8% improvement) using augmented data.
Conclusions:
- The proposed CBAM-BCEL1-INGAN method effectively augments seafloor sediment images.
- The method enhances image quality, details, and classification accuracy.
- Validated effectiveness in alleviating data scarcity for side-scan sonar seafloor sediment image analysis.
