Imaging technology advances impact marine science, providing vast underwater image datasets. However, labeling these images to identify species and particles is a real challenge. This study shows promise for effectively classifying underwater images and aiding marine science research.
Advancements in imaging technology have greatly impacted marine science, allowing researchers to gather large amounts of underwater images from different species and locations. However, before these images can be used for research, they need to be labeled to identify the species or particles they contain. Manually labeling such a large number of images is time-consuming and a bottleneck in the analysis process.
To overcome this, researchers are using deep neural networks, which can automatically classify images. They have shown promising results in accurately classifying underwater images without human intervention. But there are some challenges. Real-world performance of these classifiers doesn't match the impressive results achieved on benchmark datasets. One problem is that there are not enough labeled images available for training. Also, some species or particles are rare, making it difficult to recognize them accurately. Additionally, imperfect visibility and image focus can lead to expert disagreement in labeling the images.
To address these issues, Tobias Schanz (PhD student at Helmholtz-Center Hereon) and David Greenberg (PI at Helmholtz-Center Hereon) have conducted a study to improve classifier accuracy and efficiency by:
1. Rebalancing the data and using a Bayesian correction method to handle the imbalance in class representation, which leads to better recognition of rare species and overall accuracy.
2. Using the algorithm SimCLR to teach the neural networks to understand image features without relying on labeled data. This reduces the need for a large number of labeled images.
3. Dealing with inaccurate or conflicting labels, using a specialized method that takes into account the input from multiple domain experts. This makes the classifiers more robust and less likely to be overconfident in their predictions
The researchers tested their approach on real underwater datasets and observed significant improvements in accuracy, efficiency, and processing time. This study published in Machine Learning: Science and Technology shows promise for effectively classifying underwater images and aiding marine science research.
Original publication:
Tobias Schanz et al 2023 Mach. Learn.: Sci. Technol. 4 035007. DOI: 10.1088/2632-2153/ace417