Deep learning with self-supervision and uncertainty regularization to count fish in underwater images

Area of Science:

• Ecology
• Computer Science
• Marine Biology

Background:

• Accurate animal population monitoring is crucial for effective conservation strategies.
• Traditional methods for counting animals in the wild are often inefficient, costly, and limited in scope.
• Image-based monitoring offers a less intrusive and wider-reaching data collection method, but efficient analysis of this data is challenging, especially for densely packed or noisy images.

Purpose of the Study:

• To explore the application of deep learning for counting aquatic animals, specifically fish, in low-resolution sonar images.
• To introduce a novel deep learning framework utilizing density-based regression for efficient and accurate animal crowd counting.
• To enhance the counting model by incorporating self-supervised learning with unlabelled data and uncertainty quantification for improved biological decision-making.

Main Methods:

• Employed a deep learning, density-based regression approach to count fish in sonar images.
• Created and utilized a large dataset of sonar videos of wild Lebranche mullet (Mugil liza), including 500 labelled images.
• Implemented a self-supervised learning task using abundant unlabelled data to augment the supervised counting task.
• Integrated uncertainty quantification into the model training process to provide a measure of prediction confidence.

Main Results:

• The proposed deep learning framework demonstrated superior performance in counting fish compared to existing deep learning models.
• The model's generalisability was validated by testing it on a separate benchmark dataset (DeepFish) of high-resolution underwater images.
• The integration of uncertainty quantification improved model training and provided valuable insights into prediction reliability.
• Outperformed other deep learning models on both the sonar dataset and the DeepFish benchmark.

Conclusions:

• The developed deep learning framework offers an efficient and accurate method for crowd counting aquatic animals in visual data.
• The study provides an open-source framework and training data, facilitating further research and application in wildlife population assessment.
• This work contributes effective deep learning-based tools for monitoring natural populations, addressing the growing need for efficient analysis of large visual datasets in conservation.